The competition is real in pricing. Thanks for the Chinese open models, US big players have to cut their inference pricing. We've done a bunch of evals between the models, and Kimi K3 was the first one that actually could compete or be even better than Opus or Sol in our use cases, with a fraction of the price. All our developers use K3 as their programming model, and it now powers a big part of our systems instead of Opus and GPT. Surprisingly the new Sol pricing is quite similar to K3...
Now DeepSeek v4 Flash 0731 is eating Gemini's lunch, and suddenly we saw a price cut (the "introductory price") for 3.7. DeepSeek is of same quality or sometimes better than Gemini for text, Google knows it and they have to compete. Too bad it's too little and too late, it's still 4-5x more expensive in our evals.
And these models are not going away, nor their prices going up because of competition in the inference providers and due to the fact that you can buy/rent the hardware and run them in your own premises.
It's funny how quickly we went from "the greedy US companies are subsidizing prices to keep competitors out of the market" to "the greedy US companies are overcharging because they are greedy."
I'm not happy about it, but API prices seem actually reasonable if you compare it to free-market pure inference providers. They need to buy the same power, same hardware as the big labs without needing any R&D. This means that the coding plans must be subsidized. I'm not happy about that, because it means we'll be paying more, until hardware and maybe power drops in price, which, if we look at e.g. the housing market, may never happen during my lifetime.
No, almost everyone has been in agreement that they’re subsidizing subscriptions, but there have literally been dozens (hundreds?) of threads on HN in the past 12 months with people vehemently arguing that API prices are subsidized.
I think both can be true - they're losing money on the API and they're still charging too much. Which is really where the music stops for American AI investment. I also use K3 now and its perfectly capable for the development work I'm doing. I don't shed any tears for OpenAI or Anthropic.
API are not subsidized. We know this is true from the existence of independent inference providers (a lot of which are crypto companies that would otherwise just be mining if inference weren't actually profitable).
I don't think the rando providers like Novita, Phala, AkashML, etc., which frequently provide discounts to compete for traffic, are VC-funded. They have to actually make money to stay afloat.
Certainly not everyone. I find it far more likely that both subscriptions and API pricing are profitable in their own right. The fact that some people get great value out of their subscription (just like some people get great value out of their car insurance) doesn't mean it's subsidized.
That’s a fair distinction. I think they are willing to subsidize individual subscriptions for people who maximize usage, but their overall subscription business may be profitable.
Both can be true simultaneously. American companies were comfortable with a high cost structure because it padded their revenue numbers. The government tried to block out competitors through import/export controls, and it ended up backfiring by creating more efficient competitors. Very similar to our competition with Japan in the 80s. We'll see if we give the AI labs the kinds of sweetheart protectionism that US car makers got.
As one example: Different access patterns. If you're using the app and have a subscription you get cheap tokens in the hopes you need more at which point you pay API prices which are much more expenseive.
I agree with that. However, there’s much dialogue around subsidized tokens for business use too, that people paying for the tokens are also vastly underpaying vs the “real” cost. I certainly don’t know the answer to that. Maybe the Chinese companies are also doing it. Maybe nobody is doing it.
Looking at reserved capacity cost for PTUs on azure, which I think they’d probably not subsidize but can’t be sure, I’m inclined to not agree with the vast undercharging for tokens hypothesis.
Wait, there's 13 providers for Kimi K3 in OpenRouter. I'm having a hard time believing every single one of them provides them without any profit.
And this one is easy to calculate: take your monthly API spend to K3, then rent a stack of 8xB300 for a month and see how much it costs. I would say you're about to save 10-15k dollars per month if you have enough traffic compared to pay per token pricing.
It's not very complex math, and the hardware of course is cheaper if you bought it last year and if you have extra GPUs waiting in your warehouse (depending on if you can produce enough energy cheaply).
I don't think that kind of protectionism will work with a digital asset. It is much more difficult to erect barriers when there are no physical goods and transportation across geographic borders is instant and free.
It will likely be on the enterprise side if it happens. Hard to control supply, but if demand is concentrated to a few hundred entities, that's easy to do.
I shifted from DeepSeek v4 Flash 0731 to Gemini 3.7 flash on openrouter and price shoots up almost double with no visible change in outcome. So, today I reverted back.
I'm using hermes and keep close control over open router provider to DSv4 flash 0731
Use only the deepseek provider, you can configure openrouter to do that(but they create some obstacles, go to configurations and allow all providers)
Right now i'm testing deepseek harness, don't wait to test it. The plugin architecture and self awareness of workflows really make you think about "what is a tool vs what is a project".
Is there actually a difference in cache rates between OR and official API? I have a preset set up on OR so that I only send traffic to deepseek. The preset is important otherwise you will send traffic to different providers but if you weren't doing this already then what can I say, water is wet, of course cache rates will be awful. I get about 70% cache hit rate with the preset which is appropriate for what I'm doing. I haven't used the official API though.
I was struck by a video ad that Google released yesterday with testimonials by three developers about using Gemini 3.7 Flash [1]. The point they emphasize most is price, followed by latency. The marketing strategy definitely seems to be shifting.
Yes and no. It competes in the mid tear not in SOTA. It's a very valid model if you need things like computer use or image recognition. Especially with the 3.7 "introductory prices". It's multi-modal and better than GPT 5.6 Terra while only a bit more expensive.
And Fireworks did not yet. They are still under the limit of not feasible to self host... Let's see if other providers follow DeepSeek with their flash pricing.
Long-context agentic tasks and Rust engineering are our use cases where Kimi definitely is better than Sol. We can measure our own systems and the numbers say that Sol has no chance against K3 or Opus, and K3 is so so so much cheaper than Opus right now.
You cannot just look at the price tags for these models, you must eval and see the price per task. In our previous eval rounds Sol was more expensive than Opus (with its original price), took much longer, and provided worse results. Kimi does not have these issues, it's just as good as Opus with a smaller price tag.
China's 50 Cent Party being a real and noticeable thing (and the two biggest things they like to shill is open weight Chinese models and the futility of resisting a Taiwan invasion), I have to take things like this with a healthy dose of skepticism without corroborating data, since independent evals didn't show the price per task lead you're showing.
If there's independent data showing this feel free to share a link, I haven't seen it. DeepSWE has been most closely matching what I see in my own use.
Internal reports from company? Maybe not. I'm just saying you have to eval eval eval if you are working in this industry. There's a ton of victories in price, and price is right now the key thing all the customers are talking about.
It's not always Chinese models. For example GLM 5.2 just did not work for us at all. And Gemini is still the best cheap model for non-text agents.
If you don't have a good eval set and if you don't check the models weekly, you are missing on things. And Opus 4.8 is still the absolute quality king for agentic tasks. Too bad it's so expensive.
And the clearest thing here is that Fable, Opus, and Sol are all too expensive. I'd say a healthy 75% cut to token prices and they are back in competition.
Yep. It's a bit scary also. There's a lot of opportunity in the market now, but the downfall of the big US inference labs is going to hurt here in EU too sadly...
We know the price wars are coming. It's gonna be interesting to watch how it plays out in real time. Anthropic/Openai probably want to race to the IPO before the price wars start to have real impact.
After using Claude for a long time, I tested Sol 5.6 for the first time today. Love it, its an incredibly capable model and uses far fewer tokens/time thinking. Its what I imagine Fable would be if I haven't been downgraded on every conversation - even after completing the verification program. I think I may cancel my Claude subscription finally.
I think Fable's dominance is overstated. It definitely has the lead, but quantifying what that lead actually is is really hard. I'm using GPT 5.6 Sol to do some shit that I personally would consider "crazy" - low level undocumented hardware driver alchemy, reverse engineering highly obfuscated code, even a bit of screwing around with a rendering engine in Vulkan, really just about the most complex tasks I can get any model to do, and it does great. For the more advanced stuff, it definitely needs the effort bumped. But even with the effort bumped, the token usage really doesn't seem to skyrocket too badly until at least you hit xhigh and max, which really only seem to be necessary if you are doing genuine crazy stuff, so it's not that bad. I did similar stuff with Fable. In fact, I went directly from an Anthropic subscription with Fable to an OpenAI subscription with Sol, more or less, and it really felt pretty seamless. If anything, I was thrilled to realize how much I actually preferred Codex CLI, to the point where I started using it at work too.
Fable seems to be generally more impressive at outputting one-shot web apps. I'm not really saying that to try to downplay what Fable can do, it's just that if I compare the two, this is one of the few definitely noticeable areas that you can easily demonstrate. Obviously, one-shotting programs is much better as a demonstration of a model's capabilities than it is practically useful (not that it is useless, but hopefully my point is understood).
However, whatever Fable truly is better at, one thing I really like about GPT 5.6 Sol is even harder to quantify: taste. GPT 5.6 Sol outputs are still LLM outputs and they contain many things that people would probably consider "Claude-isms" for better or worse, but overall I really prefer the GPT 5.6 Sol output. I find it to be generally more tasteful. Hard to quantify, but when talking to people I've had enough people seemingly agree with me to convince me that it really is true.
I used Sol to extract the remaining decryption keys from the Super Mario Maker 2 (Switch) game files. Someone had previously extracted all the keys from the original release, but not any of the new ones from updates. Not only did it succeed, but it helped me understand the data sufficiently to add support for “Super World” rendering to my level viewer (which I made back in 2021), eg the little widget at the top of https://www.smm2-viewer.com/players/B16-306-GVG
I was very pleasantly surprised to find Sol wasn’t obstructive over what was clearly a very grey area endeavour.
Interesting. I've been wishing that old 'Stars!' game from the 90s would play easily on modern systems. I'd love it if we'd got to the point where I could point Codex at a folder with an ISO from my CD of the game and tell it to go reverse engineer it all for understanding of game mechanics, then go recreate it in a modern language capable of running cross-platform. Scarcely any need to improve on graphics, it could even be a PWA.
There's people that have tried to contact Jeff McBride and follow the IP trail but the IP is currently owned by a company that went defunct. Not sold, but no one is even bothering to register its LLC any more, it's simply dead.
I was having it look at creating a driver for some old scanner and it actively looked up exactly where that gray area for my country was wrt decompilation.
Fable is almost unusable for anything but super boring mainstream stuff. I was getting safeguard flagged so often I’ve significantly reduced my usage out of fear they will blacklist/ban me.
Some of the topics it’s flagged have been hard for me to understand what it seeing that can be remotely concerning in my requests.
I cancelled my Claude max subscription. Somehow every query I sent was flagged as bio or chem, even pure mathematics questions. Not going to waste money paying for a “max” subscription that won’t ever let me use the top tier model…
Sol is great and has never blocked a request, and generally gives great answers. Happily switched over to it now.
I've gotten flagged for asking questions about tokens and tensors. That makes me believe it's not about safety, it's about protecting their turf. I cancelled my subscription - same fear about getting flagged too much leading to a ban.
They said they also block usage of Claude models to build ML models.
Which is definitely protecting their turf, but also probably a little bit hiding their “RSI” abilities for competitive reasons. My theory is that a lot of “safety blocking” is actually WIP training of new business directions. Anthropic has started hiring biologists and has opened a preview of a “Claude code for bioinformatics”. I’m guessing they’re tweaking their bioinformatics market play, and block “bio safety” requests so competitors can’t learn about their training.
What's the distinction between "protecting their turf" and "competitive reasons"? I see them as the same, but I could be missing something.
That's an interesting thought on the current "safety blocking" being a trial run for the topics that scare people (bio). You're more charitable about their motives than I am, but you might be right.
My differentiation for this comment, was “preventing someone from using your product to build a competitor” vs “letting a competitor see your strengths/weakness to benchmark against you”.
Competition is competition and it’s two sides of the same coin.
I agree. It’s flagged me on discussing fast GEMM implementations for large regressions, a discussion on theoretical physics math, a discussion on designing a type of
RAG system. I’m super baffled as to what the safety instructions actually are other than “advanced anything” and even then their definition of advanced is a joke, I’m an idiot and my questions are almost laughable.
Same here - my questions were sophomore level. I think it's notable that when I edited my question to say it was about Gemma 4, it answered without blocking. A cynic like myself would interpret that as evidence they don't care about sharing information if it involves their competitors.
I feel like those examples are considered difficult because they're niche topics, but aren't actually all that difficult in a general sense. What I consider truly difficult are things like taking a ticket and implementing it in a preexisting codebase, using a clean and reasonable design that fits the existing style and makes sense to a human, and avoids the footguns I learned by working with the codebase for over a day.
If you said this in 2025 I would've 100% understood, but to be honest getting AI models to do a pretty good job on day-to-day ticket work has become so boring that we don't even bother using the top tier models and higher effort slots for that anymore. I personally wind up tweaking the results a lot and recursively having fresh agents review the diff, but that's just because I'm picky; in a lot of cases the first diff is actually pretty damn decent.
Compared to what I am doing at home experimentally, I feel like day-to-day work is absolutely nothing. Not only am I also working with existing codebases in my experimental prototyping, but I am also doing things vastly more complex with vastly harder constraints.
Personally I wouldn't want bots running autonomously on a repo, even if there were other bots cross checking them. But that having been said, I'd also say that my experience was similar with human code: it is rare to not find at least some issue worth at least pointing out. The only real difference is that the LLMs have vastly different holes than people do, making the real challenge trying to make sure you're covering them. In most cases for me the best solution seems to be just giving them a way to test and attempt to prove things out in a realistic environment. But clearly, we haven't really left the era of having humans in the loop. Fully vibe-coded codebases clearly suffer from a myriad of issues.
Multiple security holes. Issues marked fixed that aren't really. Giant holes left in solutions.
Sol over engineers now and then (hey please don't factor that function out into its own file....) but it doesn't do the same level of stupid terra does.
That said, plan with Sol, implement with terra, have Sol fix all the mistakes, then I go over the code and make recommendations for the architecture to fix Sol's foolishness.
My experience is that current models are pretty good at making functional changes without too many more bugs than a human would make, but are still bad at making good high quality changes. I have a degree in software engineering specifically, so perhaps I am overly sensitive to design issues.
I use the AGENTS.md to show it how i want the code to look like. Something like "when implementing hooks adhere to the guidelines in docs/react-hooks.md". And then react-hooks describes your heuristics and what you consider best practices. There is a clear difference in code quality for me when using codex with a well crafted AGENTS.md vs. without one, you can run the experiment yourself pretty easily. As I mentioned in another comment, I think Claude poisoned users to stop relying on their Claude.md files and new codex users might be surprised at how well it adheres to guidelines.
And Mai-Code-1.1-Flash seems like a really good cooperative player to GPT 5.6 Sol. You get Sol to help you make a detailed plan, and Mai codes it up and you can get pretty decent code out the other end without too many tokens if you are careful.
FYI I run it consistently in xhigh regardless of difficulty of the task at hand. I remember high being very fast, but I'd rather wait a bit more and get better output. AIs are insanely fast compared to me anyway, even on xhigh. Consumes more usage, but even at 100 EUR/m I don't hit limits.
After hitting the session limit on my company's plan so many times with Claude when I was using it, I mostly keep Codex on "high" rather than "xhigh" as a way to leave the tokens for my more ambitious coworkers. It's possible that having it higher might end up with better output, but so far at least I've yet to see a way to get any model to do 100% of what I need up front without any need for me to make changes that end up being more tedious to do via interaction than by hand, and it doesn't feel worth spending a bunch more tokens trying to figure out how to better communicate to it up front how the dominoes get set up so they fall in place properly the next time.
To be fair, I actually do run xhigh as my default. However, for the first time in my experience of trying and using LLMs, with Sol.. sometimes I feel confident enough to set the effort level to "Low". I just had Sol prototype some AWS stuff on low earlier. Great result, did exactly what I wanted.
Initially I was planning heavily around context limits, but I've learned to just ignore it completely. Compaction is seamless for me. If details are lost in compaction, the model just re-reads what's needed. My conclusion is that at least for Sol, the summaries (which I've never seen) must be amazing. Every now and then a detail gets lost and I have to repeat it. I don't think there is performance degration, because the model is smart enough to re-read relevant files as needed.
things that are alchemical are rarely alchemy. That is to say things are very fiddly but stick a room of monkeys on typewriters, a schizophrenic developer with HolyC and adderall or an LLM, persistence is the key to many of these things like drivers, extracting keys from vintage security domains, etc. Dropping into xdd to a human is a chore, not for an LLM.
Although I am not exactly sure what you mean, I am not really claiming it is doing anything I couldn't do - but yes, it does so with much less effort. For example, I can have it set up probes and tracing on Linux that I personally would have to consult documentation to do. It might not even have to consult the documentation due to having the information on-tap, but even if it does, it's nothing that would cause it any fatigue, it's just going to keep moving forward in a loop until it is satisfied that it meets the criteria. I could've done all of this alone - I really could have. I just would not have. Being able to do something 10 times faster or with 10 times less effort is, in some senses, sometimes more impactful than being able to do entirely new things you couldn't do before.
This is a really good point that I definitely failed to grasp when first hearing about these tools. At least for me, the best way to use these tools is as a way to free myself from having to spend time thinking about the things that aren't worthwhile so I can focus on the things that truly are. I've had times in my life spending hours reading documentation and googling random things to try to tease out the correct sequence of commands or the exact right shape of an API to be able to make things work to know that it doesn't make me more productive to do that myself rather than point an LLM at the thing and let it spit out the answer after a few minutes. Meanwhile, I can spend that time thinking about what comes next, or what the correct way to take that one-off output and abstract it to something that can be used meaningfully in more flexible ways.
The only obvious objection I can think of to this line of thinking (at least from a technical perspective) is "how does someone build up the knowledge to be able to use a tool effectively in that way if not by doing things by hand at first?" The honest answer that is "I don't know, but that's also pretty much exactly the type of thing my employers have never been paying me to solve in the first place". Even just a decade into my career, there have already been plenty of times in my career I've struggle to convince people that we should do stuff in a way that won't bite us in the ass a month or two down the line, and in the times I've managed to succeed, it's usually only by putting in more of my own time and effort to make the initial investment seem more palatable. Luckily right now I'm not in one of those times when I'm having to go full throttle to keep the lights on a few months from now, but I don't have enough fuel in reserves to work on a plan for when we need to build a new rocket in another ten years. Maybe ask me next month.
It's easy to dismiss "taste" when you either have none or just fail to appreciate it, but nothing gives you an appreciation for the importance of taste like LLMs. There is no benchmark for taste, so while many things improve taste does not. Bad taste is, in fact, a huge component of what makes AI slop so sloppy.
But human coders can have bad taste too. There is code where there is nothing obviously objectively wrong, yet the choices feel like they were made by someone who just doesn't value or put emphasis on the right things, yet spends a lot of effort on trivialities. It comes in many forms.
You are confusing personification, which is purely a rhetorical device, with anthropomorphization. I am not anthropomorphizing GPUs. There is nothing particularly weird about personifying model weights, as we do with computer programs, cars, and all other manner of inanimate objects every day.
To be honest, in this case, I wasn't even personifying them, because I in fact didn't say that an array of GPUs has "taste", or in fact even that model weights did. I was saying I preferred GPT 5.6 Sol's outputs as a matter of taste. I can see why someone would confuse the two statements since in this case they're pretty much the same thing, but if you re-read what I said I was actually more careful than you're giving me credit.
That's what they're always going to be, so not sure what would be "insane" about it. They literally feed on and emit natural language, and are put to work on informally defined, arbitrary tasks.
When people figure out any reliable strategies to test and benchmark them, that's insane, and in the positive sense. This very same issue has been a thing for humans as well forever, and remains only very questionably solved (IQ, academic tests). This is not easy.
Doesn't just sound like it, it is. That's life for you. That's what I'm pointing out to you. [0]
Just consider your own example. Do you think a less or more "rebellious look" is not something designers can actually ellicit? Less so in software design, sure, but in character design for example? Or general product design? Do you think e.g. Monster energy drinks are branded the way they are completely due to happenstance or something?
Except people don't usually put numbers to it, because they understand that that's hard to defend. You're the one who's describing such an idea, and wants such a thing to happen, classifying anything else as just vibes (that's the point!) and unhingedness. You're handwaving the difficulty and fundamentally limited nature of that, assuming that it is some laziness or mental delusion that's preventing it instead. You're also pretending as if it was somehow not real as a result. What I'm telling you is that you're wrong about that. Any kind of qualitative analysis that's actually defensible with these is genuinely difficult and limited in nature. See also all the opining about benchmaxxing. It also doesn't mean they're useless though, see also benchmarking.
The guy above didn't put numbers to his vibe assessment, they just drew a comparison, exactly because they know that there's not much else they can earnestly offer. You're sulking at them not lying to you by overstating their rigor, and you're flipping the arrow as if this limitation was some sort of mistake, not a necessary and intrinsic property, which it absolutely is. Natural language is an inherently subjective medium.
I think we’re saying the same thing?? It is literally just vibes and all this faffing over which model has better “taste” is pointless
Where you’re wrong is pretending there is any intellectual rigor to the discussion which justifies promoting from the domain of vibes to actual reasoned debate
We do agree that this is a vibes discussion, but disagree on whether that makes it a pointless discussion to have.
To give you a practical example, small, self-hostable models are very popular on HN. But my personal experience with them has been absolute dogwater, so this difference informs me that this is not the place where I should shop for a signal on whether a specific model like that is worth trying, or on whether it's game over for large models and remote models yet. I can "take the temperature" and make use of that without it having to be any rigorous, high assurance or mechanistic thing. It's suboptimal, but not useless.
Conversely, it also inspires people to try and substantiate these issues, so that it can eventually be more rigorous, higher assurance, and mechanistic. This is why I brought up benchmarks and benchmaxxing. Hard to know what points of consideration are salient when there is no abstract grievance to investigate, but Goodhart's law does also keep looming.
If people start moving away from Claude to GPT because e.g. Claude's output is too hard to work with and parse, that's relevant for the respective model providers, because it's a revenue share shift.
If I really like using Claude and strongly prefer its output stylistically, then claims otherwise will infuriate me, and will drive me to substantiate. If for no other reason, then because it's likely that Claude will have its language tuned in response to people's feedbacks and the revenue shifting, which may not be to my liking, and so I better prepare to call it and contend it.
It's literally like any other real world thing ever. Think tuning video codecs. Or tuning user journeys in frontend development.
Sol is way too eager to hone in on small details and ends up with massive over-engineering. Fable does it too - to be fair - but noticeably less.
After extensively using both on Max 20x plans, I've concluded that Fable is better for problem solving and coding, whereas Sol 5.6 Ultra shines in debugging specific issues: tackle a problem with Fable then leverage Sol to clean up, double check, or fix specific issues.
Fable (imo) had the edge on the $200 plan, but after this 50% reduction I'd say Codex is better value by far and there's no contest.
---
Using Fable as the orchestrator and delegating tasks to Sol 5.6 Ultra via the codex plugin in Claude Code yielded good results, but still there was a lot more over-engineering (thus time and tokens spent) than Fable by itself would've done.
Both models suffer from doing-too-much. But both models are fundamentally really smart and knowledgeable. I think it's really close and pricing cuts really spice things up for us consumers! Sol is a clear winner in the value department and the $100 plan is enticing!
---
*Claude Code usage is reducing by 33% in 2 days, Wednesday August 19... cmon anthropic: clau.de/cc-50-promo
Yeah, I very much agree on this. I think Sol and Fable code quality is on par. Maybe Fable is just a tiny bit better, but Sol compensates with its ability to work through things, while Fable, in my experience, generally tends to avoid solving problems that require many LOC.
However, I think these are very different models in terms of orchestration. Long-horizon tasks are way more predictable with Fable. It just doesn't lose track of details. Thus I ended up building a small wrapper around Pi (where I run Sol) so that CC can delegate via background tasks, automatically wait for completion, and do what was one of the most effective parts - steer Sol toward simplicity, getting Sol out of code-review infinite loops (Pi calls for Codex review to ship better, but generally gets stuck on P2 and results in vastly overengineered work).
One of the worst experiments was enforcing coverage at 100%. Only Sol, with an enormous amount of code and significant pushback (on architecture decisions) to Fable, was able to reach it. It made me think this is somehow related to overengineering in general, so that instructions on acceptance criteria in claude.md plus proper DX (e.g., Lefthook) actually led to okay results. It mostly helped that responsibilities were clearly split: Fable designs architecture, Sol handles coding and debugging.
I get a lot of mileage using Fable to spec, then Sol to review the spec, then Fable to plan, then Sol to review the plan, then Fable to implement, then Sol to review the implementation. It's a lot of steps, but a great boost in quality of output.
Sol 5.6 Ultra doesn't seem worth it because its long-horizon task scaling is not that good.
Remove the one thing that differentiates these 2 SOTA models (long-horizon task scaling) and you're left with assessing the raw intelligence of both models.
Both models are really fucking smart. We should be intentional when discussing effort levels when it comes so SOTA models because currently, effort levels are the essential lever to evaluate task scaling.
I've never seen that leaderboard link but I think my sentiments reflect exactly the findings: Sol 5.6 is smart, Fable is smart, but Sol is more value for the end user (even more so when you lower effort levels because that doesn't degrade the model's raw intelligence/knowledge). Not to mention Codex resets, that's just the cherry on top!
But smart != capable and this is evident once you start assessing both models on long-horizon tasks with higher effort levels. While Fable is (imo) at least a little bit better, both are still very good. If you want to test the raw intelligence of said models, you should lower the effort level.. if you want to test the model's capabilities fully, you should increase the effort level.
Economics aside, Fable is the better model (imo), but there's no need for a binary stance here. Both models are very good yet there is a clear winner on the value front.
It's great, don't get me wrong, but so is Fable. I'm just comparing the long-horizon task performance between the two at the same or similar effort levels.
Given the 50% discount on Sol and how smart it is, yeah it's unprecedented value. If you only want to use low effort, there's a clear winner here on value and it's not even close!
*The "Max" I referred to was the plan tier, not the effort level btw
For small tasks, you can just use something like low or medium effort and it can usually avoid mistakes; after all, the model will test the code anyways and can do some baseline level of iterating.
In regards to cost, we need to acknowledge how generous OpenAI was in the last couple months with Codex usage credits (no weekly limits) and usage resets. It afforded me many a dive with Codex! Yes it uses more tokens, but sometimes it's worth it -- just depends on what you're working on.
Finally, Ultra(code) isn't that bad when it comes to cached tokens. I think folks overstate the general token usage of ultra effort on both providers.
---
Both models are great at green-fielding a project when given detailed specs.
Both models overthink too liberally (imo) during these larger multi-shots. Sol overthinks more than Fable.
Both models are really smart and perform great for general knowledge and regular coding tasks.
Fable feels less cumbersome to work with, but it is SO DAMN ANNOYING with the refusals that I'm leaning more and more on Sol, and very much looking forward to GPT6. Just seems like Anthropic is trying their hardest to ruin their reputation and user experience.
Yeah I feel like I'm living in a different dimension than these people. I wonder what they're working on. I've literally never had it refuse everything and I max out my 20x plan every week
Can I ask which country you're in? I have a theory that the safeguards differ depending on the user's country.
I'm in Australia, and Fable downgrades to Opus when testing for bugs in memory in a legacy C code base. If Fable starts taking initiative and writes a test case that involves writing to a null pointer, that's the end of the conversation.
As one random example, today I had it hit a refusal loop when adding a country selector dropdown to a form, presumably because it contained a “bad” country name? I hit refusals at least 2-3 times per day, sometimes many more. The worst part is it is often right in the middle of a multi-stage task, so the only option is really to switch to Opus 5 and let it defecate its absurdly verbose comments all over the rest of the edits in the turn and hope it doesn’t go on one of its tangents, then have Sol do damage control. Oh and I got approved for their “cyber verification program” blessing, which comically does absolutely nothing for Fable.
5.6 Sol is a joy to use for "daily chat" as well. Compared to earlier OpenAI models it catches and corrects its mistakes very reliably. It also seems way smarter in tuning its replies to areas I am more/less knowledgeable about (i.e. when I ask it a law question, it assumes I know as much as a toddler which is true, but on political topics it more easily throws around terminology) and including analogies. On medium thinking, it's a very good compromise between speed and quality.
I too switched to OpenAI after I got sick of Anthropic's constant "safety" downgrades. Sol is definitely a breath of fresh air.
> even after completing the verification program
Was it easy to complete it?
I ended up in some weird state where I can't even attempt the verification at all. Opened the Persona tab once, closed it and then it never opened ever again. It says a verification precheck failed.
Even without TAC, Sol doesn't seem to get blocked very often. Fable would downgrade to Opus if I looked at it wrong.
Sol is my daily driver but there are still times I reach for Fable when Sol doesn’t cut it. Just yesterday for example, I was trying to build a self-modifying hot-reloaded agent harness in Elixir for fun and Sol just kept doing silly things like thin wrappers and unnecessary abstractions. Fable handled the task elegantly. Sol is really good as a reviewer for finding bugs due to its thoroughness however.
I cancelled my subscription recently and moved to Sol. So far - it has been a great experience. The only aspect where Fable/Claude is better I feel is doing some research from the web and summarising the facts.
I switched from claude to gpt when 5.6 came out for the same reasons. I don't understand why so many people are still using claude when GPT and open source are so much better.
Yeah at this point claude is overrated, overly expensive, weird writing style (elliptical), and the worst part is the aggressive guardrails that even normal convos get interrupted, meanwhile openAI is still I would say at the normal balance, if you ask something too obvious or direct it will stop you other than that, it work flawlessly, plus, I have yet to hit the limit despite heavily using it these past weeks.
You've replied incredulously to a similar stated experience in this thread already and proceeded to ignore the follow-up. Why are you again asking a question to which you have no intention to field an answer?
It's the complete opposite for me. The model might be the worst model I have ever used when compared to other models in the class. You just can't get it not to just write the most enterprise over complex over engineered solutions for every little thing you ask it to do.
It the first model to actually make me pissed off to use AI. I absolutely hate the model so much.
I don't even want to see the codebases this model is fucking up.
It might just be good at finding bugs that about it. That all I would ever use it for just because it works harder than Claude models.
Fable is the only one that follows my instructions correctly, which I find quite important. It takes STYLE.md and SPEC.md as law, and code just like I would code, with the same mistakes and all.
I just can't get Opus (Opus 5 is dumb as a rock, to be fair) or Sol to do that, so I exclusively use Fable for personal work. When I reach my weekly limit, usually on the last day close to the reset, I just go back to coding by hand ¯\_(ツ)_/¯
Heck, it even does security reviews and fixes, as long as I don't ask it to "attack" the codebase. I'm planning on using Kimi or GLM for that part.
The final straw for Claude was its refusal to give me a list of the most recent rapes reported by the BBC and basic information about them (location, date, names, just things reported in mainstream media). It outright REFUSED to complete this task.
I will not be told what I can and can't do by AI and I will no longer be supporting American companies run by despicable people. GPT only gets my money right now because its so fast and cheap but I'll be back to Chinese models in no time.
I have witnessed 5.6 Sol Ultra edit line after line of literally empty lines ... for hours.
I wasn't literally watching it, I came back to a goal (that it started for itself without my approval!) that had done nothing but that for some reason.
sol is much better imho than Fable but i can understand if they will perform wildly different for different people with different levels of expertise aswell as different needs. I dislike fable myself it doesnt really work for me.
Sol also doesnt _really_ work but it sort of tricks me into thinking it does more convincingly :p.
cancelled my subscriptions few days ago. (was on 100$ ones, not sure if there is diff in quality for higher tiers or not.. there might be that too).
what i hate the most is that they will make any obvious mistake you do not tell them to avoid. then on the next plan to fix it, your token limit is hit at step 4/5 -_-. Both models seem incredibly good at that mostly...
for tasks outside of coding and program design i do find them quite useful. like devops crap. maybe because i hate that, i like their help there more.
Used claude since 7/2025. Switched to codex after fable got blocked. It was still 5.5 but I knew they had to come up with something. As soon as I switched, wow. It wasn't super intelligent, but it was stable. Every day it was the same performance. This consistency is definitely worth paying for.
I have a "strategy / life-coach" project, and was surprised at how much better Sol is than Fable on it, as I've found Fable to have the edge for most things for me so far. But Sol: questions were better, insight was better, it got the brief better.
Sol has held stuff for a while to do the same sort of hazard checks I assume Fable is doing, but it always releases them. I think that's the better way to handle it rather than preventing me from seeing how far I can get generating schematics to use in Minecraft. Currently: a mostly normal voxel house.
That was an issue at least a year ago. I had signed up for a claude account on my iPhone and then wanted to sign in on my laptop but nope, not possible. Insane they still haven't fixed it.
Can somebody at Anthropic tag claude in slack or whatever goofy shit you do and ask it to add Apple OAuth to your website? Clearly humans aren't testing it.
Can't change your email address with Claude either. You have to delete your whole account and create a new one. Absolute shit software and they expect us to believe these things are super powerful world changing things.
Luna saw a huge jump after the price cut and is one of the more competitive models at the new price on openrouter.
Maybe they want to see how much market they can grab with Sol?
This might help but there are already cheaper models with Sol's intelligence more or less, the most notable being Grok 4.6 at $6/m which makes it a tougher sell
It's really only between Anthropic and OpenAI for many of my use cases, since I have a Zero Data Retention agreement with both. I'm not trusting random inference providers and especially not Elmo with sensitive data.
or Tinfoil [0]? They serve open models with container integrity attested by Nvidia/AMD enclaves. Every cloud provider offers this of course, but not usually in a way that can be shared between distrusting users for economical inference. It still relies on the open-source containers being secure, and there's probably hardware sidechannels and stuff, but personally (ie privacy not liability) I trust it more than a contract
tinfoil asks more than 10x the output cost ... $1.90 per 1M tokens instead of $0.18 per 1M tokens for my favorite model (Deepseek V4 Flash 0731) on my favorite provider (DeepInfra) currently, for example.
DigitalOcean does inference, has good uptime and offers ZDR. They are not an AI first/inference first company. They have been serving cloud products for almost 2 decades.
Batch API is a bit harder, not many models/providers support Batch. It primarily is only the Gemini/ChatGPT/Claude models that do. DigitalOcean does support a 50% discount on Batch API via them directly, not listed on OpenRouter.
I don't have experience directly with Anthropic, but I wouldn't assume anything with such high confidence without knowing the persons situation. No, an individual off the street or with an LLC and 5 employees isn't going to get a special deal from someone like Anthropic.
But if their employer is bringing millions of dollars of potential spend to the table, can tell you from years of experience that turns a lot of 'no's' to 'yes'.
I use Notion Enterprise specifically for its Zero Data Retention (ZDR) access. Fable does not offer ZDR, so Notion provides a dedicated settings toggle that disables Fable outright.
I work for a large enterprise and fable/mythos are specifically banned because of the difference in data retention agreements from those models to other Anthropic models.
The three letter agencies have and will continue make sure all of those "yes's" become "no's" again, which is why the person you're trying to claim is wrong is right and you are wrong.
You do not have ZDR with fable/mythos, nor will you ever have it again. The bedrock customers who thought they had ZDR learned rather rudely that this was not the case after the government decision that also banned non US citizens from using these models.
There's a huge jump between OpenAI, Anthropic, Google, and every other major player distilling the internet into LLMs and deliberately breaking a mutually signed contract between them and another business.
As for ZDR and court-orders, what would you rather happen there? Violate the law or comply with holding the data? I would bet that any ZDR agreement has this court-ordered risk mutually understood and agreed upon.
I see no huge jump at all. If the companies don't care about copyright law, why do you think about they care about contract law?
An LLM provider that only hosts inference and doesn't train models is at much lower risk of getting put under a court order to retain data for copyright infringement cases.
Reposting because my other comment was incorrectly flagged. What im not allowed to speak out against my own disenfranchisemnet? Smfh
I don't care how capable or how cheap Grok is, I refuse to financially support a company owned by a white supremacist that is actively working to disenfranchise me and millions of my fellow citizens.
Always has been. My prediction is that both OpenAI and Claude will go bust unless they deliver a killer product. And unlike scrappy startups, they have a pretty serious deadline because creditors will come a-knockin'.
There's little to no functional difference between Kimi, Qwen, Sol, Opus, etc. All flagship models are within like 1-5% of each other and the real moat will be what's always been the hard part: making a good product.
> All flagship models are within like 1-5% of each other
Don't know about that.
I'm using code review of my lone lisp project as a benchmark. It's a massive parallel code review where a coordinator cuts up the codebase into sections and dispatches agents to consider each part from different perspectives like quality, maintainability, consistency, correctness, rigor, etc.
Ran a complete Fable/max code review. Took over a month on a subscription. Now I've switched to OpenAI and am repeating the exact same review with Sol/max.
It's still not done yet but preliminary findings suggest Sol can only reproduce 70-90% of Fable's findings. So I think these models aren't as close as we've been led to believe.
Absolutely. I have an independent audit pass to verify claims.
My methodology consists of launching a 242 cell parallel code review matrix and committing all Fable/Sol max effort agent prompts and their full reports to a private orphan branch on the repository. This is the part that is taking me months to complete. This thing can kill my $100 subscription in about 12 hours.
When done, these raw findings will be semantically deduplicated and merged into a list of findings per model. This list will then be audited for hallucinated or otherwise made up findings. This will refine the list, and hallucination rate is its own data point. I'm also counting things like cybersecurity refusals and downgrades.
When all this is done, I'll analyse the final results and publish them on my website.
Some preliminary analysis:
Which code review lenses were the most valuable, where value is defined as number of serious issues identified? Rigor, followed by tests, robustness, correctness, and so on. I was able to create a tier list of reviewer personas using evidence! I can now run focused code reviews using the highest value lenses.
What's the most expensive code review? Correctness and rigor, of the lone lisp machine specifically.
API costs per finding? $0.91 to $3.77. API costs per serious finding? $8.94 to $28.51. All Fable.
How long did it take? 28.2 calendar days, 66.3 agent-hours.
Is it worth it to run the code review multiple times? If a review matrix's defect capture probability is 57%, then a second run captures 81% of the estimated/extrapolated defect population, a third run captures 92%, a fourth run captures 97%, and further runs yield severely diminishing returns. Probably worth it to code review important stuff three times.
What's the impact and cost of the safety classifier? Out of the 52 Fable review cells that triggered the safety classifier, 35 died without producing any output whatsoever, so 67.3% of the cells were a complete waste of tokens. 25% produced at least some output.
Does the safety classifier trigger most often on the important code that actually needs SOTA models? For the most part, yes. Fable was most often barred from reviewing the most important and complex files in the codebase, such as the virtual machine, the parser and I/O layer. These files also have the most CRITICAL+HIGH severity findings. Only a couple outliers broke this pattern.
It's not hard to sell a dollar for 50 cents. Imo, these businesses are pretty clearly not doing well financially (the revolving door of unvested C-levels is a good hint, the constant postponing of S-1s is another).
That graph looks a lot less impressive when not "annualized". Annualized revenue can be gamed in a number of ways, and is a big reason companies tend to compare YoY to investors once they're public.
Their quarterly revenue looks exactly the same. It's literally the same data. Annualized revenue can be gamed when it is fluctuating. It can't really be gamed when it is growing exponentially for three years straight. There's nothing to cherry-pick.
The Chinese models are cheap because no one is using them. But they can't actually afford (or have capacity) to serve enough people to kill the giants. This is evidenced by them all recently hiking prices or limiting usage.
It's possible that they build out in China at an unreal pace, China doesn't have concept of "community input" to drag down state projects, but then you are left giving your IP to China. Just ask western hardware businesses how well that goes.
No one is in a better position to more efficiently use it - that's what happens when you poach every top 0.01% engineer/researcher in AI.
They're guaranteed to get over whatever hump you think they're in unironically. Uber/Tesla have been in far worse situations and despite Elon being an idiot/liar you see how they performed when even the most bullish of investors called for their heads
I think I generally agree with you - OpenAI and anthropic will probably succeed here. If the ai bubble pops, they’ll come out on top.
I just think the whole “moat” discourse is silly. OpenAI and Anthropic’s success depends on the same thing every business’s success depends on: their customer base, and their continued delivery of services their customers want to pay for. Not their tech or their compute or anything else. They have no moat because moats aren’t a thing.
There is a massive difference even between Opus and Fable, same provider, before various harnesses and other optimizations come into play. Don't be deceived by rankings and benchmarks, try for yourself.
The problem is that most of the volume doesn't come from proprietary products, it comes from API use which has no stickiness.
Claude already has a killer product (claude.ai/chat is a Swiss army knife) but just relying on people typing stuff into chat is not enough to sustain the company.
The other strategy is entrenching yourself as the LLM of choice into existing products (like ChatGPT is on Apple products).
But who can take a benchmarking website seriously when they literally change their benchmarks to appease anthropic, like artificialanalysis.ai did the other day?
I don’t think that’s true. OpenAI docs don’t have this price change. I assume they’d be the source for this post if it was true. The banner on OpenRouter for me says Gemini 3.7 discounted for a limited time, but if I click through that I get to this page: https://openrouter.ai/models?discount=true
That shows a bunch of models, including Sol, with a discount. None of them say how long it’s for, but I’d assume in all their cases it’s for a limited time as the banner said, and only on OpenRouter.
As I understand OAI is offering discount only for users using the model via OR. Not sure why. Maybe they want OR users to try the model and switch to OAI subscription or something.
It might also be worth noting that the discount is only if OpenAI itself is used as the provider via OpenRouter. The discount does not exist for Azure or Amazon hosted Sol.
I posted elsewhere, but the Azure uptime & performance for Sol is truly dire. OpenAI is offering 5x faster latency, 4x faster tokens generation, and vastly better uptime (Azure US has only 87% uptime), all for 50% of the price now. I assume the pricing is to compete with other shiny new models (Grok, Qwen etc), but it might also be to cut-off a truly poorly performing Microsoft hosting experience.
I have switched to Chagpt sub now after only using Claude for coding. You get more value for your money and feels like codex has reached Claude code performance in coding (the reason for using Claude) regular plus account allows you to have access to their most powerful model, image generation and asking questions is better because you can use sol but in instant mode and it feels smarter and faster.
And finally codex usage limits are better than the Claude daily 5h limit.
And codex feels faster although Claude code had more features
I switched as I felt Codex was on a par with Opus, but the chat responses from Sol are just more intelligible than the word soup I've been getting from Opus. I wonder if Opus could be prompted to respond in simpler prose via agents.md
I have also noticed that I am increasingly struggling to read what LLMs are writing, finding it incomprehensible half the time.
I stole Matt Pocock's line of "When reporting information to me, be extremely concise and sacrifice grammar for the sake of concision." for the agents.md
It makes it a little better. I also specify to use https://github.com/AminBlg/SimpleEnglish/ for all writing it does including code comments. It all feels like a bandaids but that seems to be the best we can do right now.
Ah, so I am not the only one struggling with deciphering Opus writing style. At times I feel dumb as a rock because I read the same passage like 5 times and I still don't get it.
Yes it can, it's called "output style" under `/config`:
```
This changes how Claude Code communicates with you
1. Default Claude completes coding tasks efficiently and provides concise responses
2. Proactive Claude executes immediately, minimizes interruptions, and prefers action over planning
3. Explanatory Claude explains its implementation choices and codebase patterns
4. Learning Claude pauses and asks you to write small pieces of code for hands-on practice
```
You can set up a custom one under `vi ~/.claude/output-styles/eli5.md` and `eli5` would show up in the list above:
```
---
name: ELI5
description: keep it simple please
keep-coding-instructions: true
---
I can't process lengthy text, talk to me like I'm 5.
Small words, short sentences, short paragraphs. If you have to use a big word, explain it right after. Only return what's actually necessary. Just tell me what you did, did it work, what do I do now.
If I have to decide something: show 3 options max, the context I need to pick fast, and which one you'd go with.
Keep paths and commands exact. I have no brain cells left for the rest.
```
I'm not sure if it's a good thing that there's a config option for this, it kinda means they know there's issues with verbosity.
> I wonder if Opus could be prompted to respond in simpler prose via agents.md
God knows I've tried. I've got a variant of the ASD-STE100 trick which does the job, mostly, at the start … but get to about 100k of context and it goes out the window.
The model's personality is too strong for simple suggestion, alas.
I used over a billion tokens per day of gpt-5.6 sol xhigh starting last Wednesday through Sunday before reaching my reset limit. The $200 pro plan is still the best deal.
There's no universe in which just buying a second or a larger subscription isn't a billion times better and cheaper than any kind of comparable local workflow.
Privacy, experimenting with ML and "unorthodox" needs are currently the only acceptable reasons to do local.
Well I wouldn't say a billion times better. I've actually been having a surprising amount of success working with local models. And my investment has only been the equivalent of 4 months of a Max x20 subscription.
Experimentation and privacy are definitely advantages, but it's also quite a lot of fun.
Yeah, it sure was convenient that there was a RAM pricing crisis right when Apple was making local inference viable. All because of a promise that AI companies will buy more of it... with money they don't yet have, whereas Apple does have lots of money.
Agreed I’ve seen what they’re paying at work for the OpenAI API but I also think that includes reserved capacity and ZDR O_o. Try some add on credits next time if you can. I was curious at how far $20 would go (500 credits). Watched them go to zero over an hour and assumed it’d stop. It then ran for another 6 hours and completed the task despite the meter at 0. What a task means is very unclear but it’s definitely not pricing sol at $20/500 credits an hr in tokens.
I mean this is kind of standard in a lot of industries? Businesses get reliability/support usually for the price increase. Just look at lots of industrial tech. Or business vs consumer 3D printers, or business vs consumer laptops etc....
Pretty basic. The codex app with one conversation per project and several running simultaneously all hours. I’m going for max caching that way and it never gets lost even with compaction somehow. Each has a plan with milestones to keep up to date and a thin agents file. I check in on them in the Remote app. Use case is protocol and control reverse engineering of audio hardware. I think they must be identifying the heavy use agent sessions and cranking up their cache lives so it’s not a big deal for them.
Ultra mode spins up many sub-agents. On a particularly challenging task, I’ve had as many as 29 agents working at one time.
Also if you don’t specify, most end up being the same as the parent model which is pretty wasteful.
I engineered a skill that spins up Terra High agents for most sub-agents, resorting to Sol Medium for technical research and Luna High for code/in-project research tasks.
On a slightly different topic, Luna Max is incredibly capable and doesn’t use as much quota (Luna tokens are dirt cheap).
Most importantly, are you seeing a return on investment for time and ultimate outcome?
No one can judge the enjoyment, learning, and hobby aspects. Just wondering if there is an end goal for that much overall expenditure (time, money, energy, etc.)
I absolutely see a return. I do not pay for this (we have a corporate gateway), for the price of a junior developer I can get 3-4-5 senior developer's work done. it's insane value.
I'd like to see a benchmark on this specific topic: Reverse engineer the hardware protocol from a driver, or just migrate a driver from one OS to another.
I’ve chipped on it with each model since 5.2 but 5.6 sol is something else. When it first came out I’d get some refusals but they’ve since stopped. I wonder what an ideal candidate benchmark task would be for that?
5.6 has been a huge pivotal change in reverse engineering tasks for me too (largely extracting game assets from binary client files). Something I spent literally weeks on in January with claude models at the time was solved in about 30 minutes with 5.6 Sol at medium just yesterday. It's both extremely satisfying but at the same time also a little annoying how much time I had previously spent on it only for it to be solved so quickly now. I suspect improvements in AI will continue this happy but annoyed trend.
Some people just do crazy stuff. For example this now ex yc guy who said he has agents constantly scanning Sf govt apis and forming dashboards just because
Yeah I am, building large software with a vision - requirements - architecture - plan - code workflow. One Claude max account is enough to work on one, maybe two of those at a time (call it 15B tokens/month per project)
I’m exclusively using ultra and I run out in 3-4 days consistently. Those resets are great but I’ve noticed they like to cluster them at the start of the cycle, would be better if they spaced them out more.
A billion tokens per day?? Plausible estimates put the energy use at about 0.001 Wh/token, which means you're using 1000 kWh/day in electricity, just to generate slop. That's about the same as 50-100 houses. 300kg of CO2 per day - roughly the same as flying from London to New York every three days.
I think on average AI energy usage is not as big a deal as everyone is panicking about, but your usage is truly absurd and I don't know how you can live with that. It's immoral.
I can't speak for that guy, but I'm a physicist and work in clean energy... So it's not too hard! That said, I usually am closer to 10M on days I do heavy coding, so not nearly that bad.
Have you seen the average physicist code? I've crawled around in the muck of Fortran codes written in 1970 with unreadable 8 character variable names (the first of which encodes the type), with implicit typing, zero comments, computed go-tos, SAVE, the works. Tests? Never heard of it. Bugs everywhere. You would be far less sanguine about physicists coding if you looked into some of the "flagship" codes used today. And believe me, it isn't a "boomer bad" phenomenon. Every incoming student I've had in the last decade writes shit code too.
My (only somewhat facetious) opinion is that physicist access to programming languages should be controlled like doctors' access to opiates.
0. It's not about "looking pretty", it's about being nigh impossible to understand or reason about. There is a 1000 LoC subroutine that I'm convinced was ported from FORTRAN 1 unchanged that cannot be understood by anyone under the age of 50 that underpins one of the "flagship" physics codes in my field. It's a fucking nightmare every time I have to worry about that stupid thing. I remind you that this was 100% the product of a human brain (pre computer age, pre social media, and I'm told quite the genius too). Perhaps one should not place the human brain on too high a pedestal and deify it.
1. Why did you take from my OP that I tell codex "write a climate simulation code, make no mistakes" and go suntanning on a beach in the tropics for the rest of the semester?
2. Perhaps you have a different experience from me in writing HPC codes, but my experience is that > 90% of the code is boilerplate. I find GPT 5.6 can be prone to overengineering, but with a little steering and good judgement it generates very nice interfaces and high level code. I just have to think about the solver structure or metastructure.
3. Even with core numerics - pre-AI, it was a bunch of iteration going back and forth between code and optreports. Now codex will just do it. I suppose this may seem grim to you if you loved decorating every variable with !DIR$ ASSUME_ALIGNED, and manually batching array operations or whatever, but I didn't and I'm glad I no longer need to.
4. I'm now highly motivated to write tests, and AI makes it way easier to write the immense boilerplate around good tests (sorry not sorry, my {FUNDING_AGENCY} program manager doesn't give a flying fuck what my test coverage is, and my next grant won't depend on that in the slightest, so pre-AI I did the bare minimum. You can argue that the results will be worse, yadda yadda, but the incentive structure that {FUNDING_AGENCY} has in place don't promote good software standards, and my career never suffered for it)
5. I can generate docstrings with high accuracy (see the above)
As a company... but that includes things like research costs, model training etc. to determine if they're selling electricity at a loss you should look at inference costs bc that's the "thing" they're selling
I can’t sign up for that. I tried authorizing Codex a couple days ago. For some reason, their system says my phone number has been used for verification 3 times even though it definitely has not. I’ve had this phone number for over 20 years. OpenAI support is useless. They just keep repeating the policy without actually helping me.
Use TextVerified, load up like $5 of credit and OAI verification is like $1.00. Then when your account is made, ensure 2FA/passkey is setup then you don't need to worry about the phone number.
Historically those are less useful because some of the verification systems require a real phone number and that your name is associated with the account, depending on what and how they verify. It's annoying, I use a google voice number as my primary, and it often gets rejected.
There are several tiers to these services, some are selling real us phone number verifications at about 0.5usd/text while others are selling virtual phone number verifications at much cheaper. From some limited experience with the former, there is rarely if ever any problems with rejections.
Yes I filed a support ticket with them and explained that their system is broken and they just did not care. I explained how it was impossible for me to use it 3 times already as I've only made 2 chatgpt accounts EVER, and only recalling entering my phone number for one of the two chatgpt accounts. I told them that this issue locked me out of codex and chatgpt for work and they weren't willing to do anything about it. Totally useless support.
I ended up borrowing my gf's phone number just so I could get access for work. Ridiculous
Or it's "normal" market segmentation. OpenRouter users are more price sensitive in general, also a lot of enterprise users who can't switch easily are using the official API (or Bedrock or Azure) and you want to squeeze them as much as you can.
A surprising amount of companies sell exact the same product through different channels for different prices. A good example is Apple.
Multiple times a year, retailers here in Australia have co-ordinated sales on Apple products. Apple.com or their retail stores don't have these sales.
But they're clearly Apple-funded when competing retailers launch the same sales on the same days; and the margins aren't enough for retailers to take a loss.
Seems reasonably clear to me? Potentially bringing in more customers who use OpenRouter for trying out all the models with rapid switching, enticing them to use Sol. And anyone being routed on price will immediately be switched to OpenAI's servers instead.
It also seems to be providing a vastly better user experience - Azure has less than 99% uptime (Azure USA only has 87% uptime), latency of 20 - 30 seconds, and a mere 8 tokens per second. OpenAI is offering 32 tokens per second (4x faster), 4 seconds latency (5x faster), and all for half the price of what Microsoft is charging for a vastly inferior experience.
I think being able to A/B test price is invaluable for them. They can't cut prices by 50% and hike it again. By letting others slash the price, they can tell if it is worth it or not to do this officially.
if I was going to IPO I'd want higher marketshare and openrouter is the most popular way to measure model usage (not saying it's the most accurate b/c it's not)
If they can cut the price of Sol by 50% and the price of Luna by 80%, then the original price might have carried a massive operating margin. They might still be serving the models at a profit after these price cuts, but we will never know.
I don’t think there’s a real answer for this. Margin depends on whatever number the accounting department wants to make up.
Do you include research and training costs? Of all models or only the ones being served? What percent of the R&D budget do you allocate to inference? What about data center capacity? Do you count future commitments? All the circular financing deals? Do you count employee equity grants as costs? At what valuation?
We also have another solution for "whatever accounting decides": generally accepted accounting practices. It's far from perfect, but GAAP figures are what you should be looking at; not "adjusted GAAP" or whatever invention.
It is a well-accepted fact new shoes make you faster. Current science suggests it is due to the lighter weight from lack of dirt, though there is a competing theory which says it's an optical illusion due to the fact the pure white streaks resemble speedforce.
I would in such a scenario expect the GPUs to be dumped to industrial breakers who would send them to China for refurbishment and repackaging before being sold again on Amazon, AliExpress, and Taobao as last gen gaming cards from weird brands and specs.
This is what happened after the great crypto GPU dumping.
The e-waste recyclers are pretty low on the pecking order, as the creditors will be first to strip these places for assets as Leopold Aschenbrenner discovered. =3
It's where the stuff ultimately goes since the creditors aren't interested in GPUs that no longer have much value. The context here is a crash, not just a basic bankruptcy. If it were that then absolutely it would be impounded by creditors (virtually or in reality) and then sold to the highest bidding data center.
Yeah, I ment it as a joke - I agree with you. Watched the Gamers Nexus GPU investigation recently, where they were shown how a chinese soldering shop can transplant GPU chips to a new board, including memory chip reuse.
Hopefully we can look forward to all that useless datacenter AI crap gets repurposed in a similar manner into something actually useful for users.
Since it's only discounted on the standard "OpenAI," non-ZDR route (old pricing on Azure), I'm guessing a lot of users won't see this benefit? Since a lot of users enable a global "ZDR-only" toggle on OR
I'm not having that experience. So far each major model update has been at least slightly better than the last, in ways I've found useful. Can't say it's perfect, or able to do exactly what I want without a decent amount of instruction/implementation/docs, but it's been useful enough to keep paying for it.
Oh no the models are absolutely getting better, I'm just amazed that only 6 months ago I was using gpt-5.3-codex, and now I can use gpt-5.6-luna for similar results at like 1/15th the cost. Now 5.6-sol is being slashed by 50%? Amazing.
I like to see this. I still prefer Fable (marginally) but my last big task was 100% Codex using Sol max (re-sizing my AWS infrastructure using CDK) and it did a very good job. No complaints, I could use this model happily to do what I need to get done.
If this nudges Anthropic to give me more Fable usage, that's even better.
> my last big task was 100% Codex using Sol max (re-sizing my AWS infrastructure using CDK)
Fwiw, you could do this with any small or medium model, and it's easier with the aws-docs mcp. AWS is pretty stable, well documented, and programmatic, so most AI can figure out what it needs pretty quick
I'm not claiming an eval on what is or isn't possible. Rather I am stating my satisfaction with what I actually used and the actual result I obtained.
What I can say is that over the course of ~1 week I was able to review, plan, implement, test and release a significant change to a production system using Code Sol max. Any other claim about how any other model might have completed the same task is outside of my experience.
I doubt bots, the linked page is really boring to read and looks like a dashboard instead of easily consumed reading. I think people are therefore drawing conclusions from the headline more than usual since the link makes the real information sort of opaque.
However it's ignorant to think that there aren't bots on HN and especially for the very many motives people have for swaying public opinion.
You can train a LLM to inverse summarised thinking into thinking text. It’s not perfect, but it gets you maybe 80% of the quality with proper techniques.
FWIW, there’s not that much value protected here anyway IMHO, and even raw thinking text can lie (as shown by Anthropic’s amazing research), so for legitimate interpretability research it’s limited.
Scaling frontier performance hasn’t been SFT-bounded for a while now; it’s now basically how much you can scale RL rollouts.
Has anyone had mixed experience running Ultra with and without /goal?
I come back to it after 8 hours to find it got stuck navel gazing imagined and Byzantine errors.
OpenRouter doesn't do sales, they charge a premium, which is a flat 5.5% taken out of your credits. If you see something cheap on OpenRouter, it's because that one provider lowered its price. (Actually, correction, they will take 0.5% off their fee if you allow them to train your content)
Another thing some people don't notice is flex pricing, which is way lower than default pricing, for slightly worse latency and reliability. Depends on the provider and model
OpenAI is making some really boneheaded moves these days.
They're reacting instead of leading, basically.
Cutting API prices 50% while millions of your paying subscribers have had their limits slashed and are all literally looking at the salivatingly-cheap chinese API prices availalbe on openrouter...
Not only did OpenAI and all of their cash somehow MISS the opportunity to purchase OpenRouter ...
Now they're giving a discount on an API that nobody even uses (get real, nobody's paying API prices to OpenAI ...
I calculated a 5.6 sol coding session the other day ... $680+ USD ... and it actually destroyed the codebase it was working on during that session).
Needless to say, I will not be spending another dime with Codex or OpenAI.
This entire Codex reset limit fiasco has taught me they are not to be trusted.
That's some bonehead take. For very starter, neither the other AI companies nor the customers will trust OpenRouter if it's owned by OpenA. It'd be squeezed to death from both sides. The only reasonable way for OpenAI's investors to have a share of OpenRouter is to invest directly, not via OpenAI.
> I calculated a 5.6 sol coding session the other day ... $680+ USD ... and it actually destroyed the codebase it was working on during that session).
Yeah, sorry, skill issue. If you let AI run wild (if one session is $680 yeah it ran pretty wild) don't complain how it messed up your codebase.
>> Yeah, sorry, skill issue. If you let AI run wild (if one session is $680 yeah it ran pretty wild) don't complain how it messed up your codebase.
Not the OP, but I consider myself a skilled and heavy AI user, with multiple subscriptions in both platforms, plus OpenRouter.
A couple of weeks ago I gave 5.6 Sol a small/medium sized ticket to simplify part of the auth system. The ticket had a lot of details another Sol agent had collected during an exploratory session, and it was all vetted by Opus 5. I thought to myself that the implementation agent should have everything it needs. I still had it write a plan just in case, read the plan, made sure it matched the ticket, then clicked Approve and walked away.
I came back later that afternoon to a horror show. The agent had written 25,000+ LoC in the worktree. After 15 minutes of skimming through it, I realized that it had made the specced change, then convinced itself that it needed stronger verification, and over a series of compaction cycles ended up writing a static analysis harness so that it could prove that the change would be safe. Total bonkers.
Except, according to another Sol agent I showed the worktree to, the harness didn't actually do what the original agent claimed. The review agent said 98% of the worktree's code should be thrown away, and only the fix and its relevant unit and integration tests should be retained. I also asked Opus 5, and it theorized that 5.6 Sol must have gone through too many compaction cycles and lost track of its original goal.
This never happens to me with Claude models. Yes they write a lot of code and verbose comments, but I've never had a situation where a ticket that should take several hundred LoCs ended up with tens of thousands. When Claude overengineers something, I catch it during the planning phase, and it implements plans faithfully.
5.6 Sol is simply unreliable. It's too relentless and doesn't know when to stop. That's probably what caused the OP's $680 incident. I find it fascinating that people like it so much.
Is this about Kimi or about Claude? Anthropic is circulating a $200B 2028 revenue target ahead of their IPO. If OpenAI plans to stay private longer, which their recent liquidity event might suggest, why not try and kneecap their competition?
This is the opening salvos of an all out token price war.
With models a commodity at this point there isn’t much leverage for the big labs to keep their pricing anywhere near where it’s at. And that’s at the worst possible time as they need to be dramatically raising prices to have a viable business model.
Expect pricing to rapidly fall towards the underlying cost of compute and as players get really desperate we’ll likely see inference at less than the cost of compute as the market starts to rationalize and squeeze out weaker players who’s only play left will be to be the cheapest option in town.
The AI bubble is just waiting for the first player to scream mercy and cut capex as they simply can’t afford to throw more cash on the burning pile. That will be the trigger that implodes this bubble.
> The AI bubble is just waiting for the first player to scream mercy and cut capex as they simply can’t afford to throw more cash on the burning pile. That will be the trigger that implodes this bubble
Deepseek v4 and Kimi k3 have tightened the screws on the frontier models. With open-weights, anyone can host these for cost of compute. So, there is zero leverage left for the frontier labs.
What implodes it in my view is that enterprise adoption will stall. Enterprises are struggling to actually use these things in real workflows outside of coding and support.
I saw this for Luna and then looked at the uptime and it said 85%. My interpretation is that this is just a gimmick where they serve the OpenAI flex tier at the same discount OpenAI provides for flex and then fall back to azure
I think what is going on here is that OpenAI are employing price differentiation to capture more of the market. The captive API customers are already there. There are bunch of potential customers that are price sensitive and are at openrouter. This move allows them to capture more of the market and increase profits (yes, they might be increasing profits at this level)
One of the selling points from openrouter is exactly that though, it makes it extremely easy to jump from model to model. For some background tasks for example I setup all the free models on openrouter and if one is not available it jumps to the next one.
Perhaps not most but surely some, and in the future perhaps increasingly easily.
I’d have thought that even today people would validate a number of models for certain tasks and on a daily basis go for the cheapest provider when they run that task.
To me, it looks like the leading labs are investing more and more to improve their frontier models, and are being forced to charge less and less for them due to competitive pressure.
Maybe I'm wrong, but "reasoning as a service" is looking more and more like a... commodity.
That suggests this is being funded by OpenRouter, and there's no indication this is the case (and I doubt OpenRouter can afford it; why would they anyway).
Reading the comments in this thread, i honestly dont get it. 5.6-sol has felt like a regression in capability. In fact, every model since 5.3-codex has been a regression from OpenAI. I just find 5.6-Sol over engineers problems, takes absolutely ages to solve basic problems....
At this point, I'm considering going back to cursor over codex due to the ability to get more control over what model I use since there is clearly a heap of user preference and having frontier providers constantly shift the goal post with "State of the Art" is complete non-sense.
It depends on what effort you're using etc. As an example [1] of what codex is capable of, here's hugo (written in golang) ported to TypeScript - and then a TypeScript to Rust transpiler which converts arbitrary TypeScript into Rust.
The TypeScript code which was transpiled into Rust (and is compatible with most hugo templates) runs faster than the original hugo.
I find all effort levels of sol are the same in terms of amount of hallucinated unnecessary changes. Luna is much better all round on xhigh but my point still stands, every release of these new models is not an upgrade, its re-learning how to work with it.
Its like rehiring an employee every few months then training them up. Its honestly tiring and cant stay like this.
That's not been my experience. My prompting methods haven't changed much between recent GPT releases. I do put a lot of effort into building tooling and tests around a project, so the LLM output is converging around it.
They were built specifically for testing the C# target. There are several other large projects we built specifically for e2e testing.
But more interesting would be the tooling built to support this. For example, our current TypeScript parser [1] is a file-by-file port of Microsoft's TypeScript V7 compiler written in golang. The challenge here is that every time Microsoft changes code, we'll have to fix our code and tests. It's doable, but a fair amount of work.
So we decided to write tooling to transpile Microsoft's v7 compiler from golang, and autogenerate our compiler. That tool is called gotots [2] - and it already produces a fully working TypeScript compiler. It's 3x slower than TypeScript v6 compiler, but we hope to get to rough performance parity in a week or so. Everytime Microsoft makes an update, we run gotots and our parser gets updated as well.
My general point is that tests and tooling is tremendous value, and they are guardrails for LLMs to converge. I could have, for example, chosen not to write the go-to-ts transpiler, and live with porting Microsoft's parser line by line. But making such tools is something LLMs are good at, so it's a tradeoff well worth making. And the upside is that you don't have to use LLMs to port Microsoft's parser/compiler (a large and complex project) line by line.
I mean it added additional changes when it doesnt need to. Its basically hallucinating changes it thinks it needs to make regardless of effort levels i try.
No, this is OpenAI doing the discount, not Openrouter by themselves. OpenAI is crushing it with their 5.6 models, and they probably decided there was no better time to grab as much market share as possible.
I don’t understand this at all. They have never been profitable yet. How is this helping them? When it be more likely the case that not enough, people are using it as the prices they established already? So now they have to lower the prices?
You lower prices for marketshare. Fable became a mythological model to leadership because they were the first story of ai escaping and hacking another company. The it's so dangerous the public can't use it narrative is sticky so OpenAI is showing off its model so as many eyeballs as possible. We're in the samples in the supermarket phase.
Not sure as OpenAI models (Sol, Luna,..) are also discounted on the Vercel AI Gateway rn. My bet is on OpenAI trying to drive more enterprise customers to their models through API.
In my experience, "Sol" stands for "Stupid overengineering LLM". I’ve tried it at low/medium/high/xhigh effort levels and after a while I always end up to regretting my switch from Opus/Fable.
These threads seem to have become exceedingly vibes-based.
Yes, something may now be cheaper or more expensive or whatever, but there is no way to objectively measure quality (except for "trust me bro" benchmarks). So the discourse is people saying that for them, this or that model was better - which is a very low value data point.
Yes, programming language discussions can be vibes-based and therefore low value too, but sometimes the more concrete aspects of the languages at hand, e.g. language features and tradeoffs are discussed. That is something that I'm not seeing in equivalent AI discussions.. there is a lot of how a particular AI model "feels" to interact with.
It's got to be because AI has become a commodity, people are looking to get the best value for money, and things change based on your specific application (which may be nearly unique) and literally the time of day.
It's a bunch of bread bakers talking about wheat suppliers.
This could cause the AI market to crash. If companies can't make money off their users, their entire financial plan will collapse and their stock prices will plummet.
This is honestly impressive. Cutting the price by 50% while pushing a model this capable is exactly the kind of move that makes advanced AI feel genuinely accessible.
I stopped using superpowers because it wanted to turn every tiny bug fix into a $37MM DOD project. I got effective results but it took ages. I may try again - I need to find a good way to run different profiles in my harness so I can easily shut it off. The default planning workflow in OMP is pretty good though.
I agree Luna is great for task execution, either as a sub-agent with Sol planning and coordinating or if the task is well defined and straightforward, but there are lots of models now that you can say that about.
You would probably get better results with Luna for the real simple tasks, or Sol with low thinking effort.
I find that I get exactly the effort that I asked for, which is pretty nice. The other side of that coin is that these are the least lazy models I’ve used so far. They will go on elaborate tangents to complete the task when I want them to.
I've found it's worse for simple tasks too, and I have to give it stricter guidelines, and sometimes it doesn't follow the same patterns I've grown to expect. I've found using 5.6 (sol) is good for diagnosing issues though, especially in terms of optimization of some given path
Yep. I have not yet had a single good experience with Sol or the 5.6 models on a variety of harnesses and configurations. It overthinks, overcomplicates and often makes my code into an unmaintainable sludge. It'll usually take 5+ turns of steering to get it in the right direction.
It's your responsibility to set an appropriate level of Thinking. For simple tasks, I use the instant model. As an approximation, the choice is proportional to the amount of time I want it spending on the task. Also, you can always ask it to respond succinctly.
It shocks me how little people seem to care that they are supporting an evil Zionist lizard man who molested his sister and is happy supporting trump. Doesn’t even come up in the conversation here. I don’t really care if sol is a bit better, I still make decisions on more than that.
Is the HN community just too online and sucked in to the musk mind manipulation vortex? Or what is going on? Why does nobody seem to care?
I’ve used Claude exclusively for the past few months
Was excited when Sol came out a few weeks ago and loaded it up
I made the mistake of treating it as if it were Claude - I’d assumed they were close enough in ability and treated them that way
Well, turns out my instruction sets for Claude are 100% too complicated for Sol
Sol made the stupidest assumptions, constantly did things that it wasn’t asked to do and always approached code in what I considered a weird way - I had redo a lot of my prompts to get it anywhere close
Now, did it do good work?
Yes, on occasion. But with LLMs and coding, consistency is the name of the game. Constantly having to correct the LLM and constantly feeling paranoid that it won’t listen makes for an exhausting session
Maybe if you “came up” in the codex world you’re more fluent with it, but sticking with Claude for now
My bet is that OpenRouter began steering GPT-5.6-sol users towards flex tier, which is already 50% off.
So this isn’t really a price cut. As to why, lots of possible reasons. Perhaps an agreement with OpenAI to help them drive up more diverse traffic priorities.
Now DeepSeek v4 Flash 0731 is eating Gemini's lunch, and suddenly we saw a price cut (the "introductory price") for 3.7. DeepSeek is of same quality or sometimes better than Gemini for text, Google knows it and they have to compete. Too bad it's too little and too late, it's still 4-5x more expensive in our evals.
And these models are not going away, nor their prices going up because of competition in the inference providers and due to the fact that you can buy/rent the hardware and run them in your own premises.
reply