Forget about the scores... look at the Cost Per Task. Fable 5.1 was already the most expensive by a good margin, and now Fable 5.2 doubles it! This is with the price cut.
The output tokens per task just exploded. From 118k tokens (46k Answer/71k Reasoning) in 5.1 to 274k tokens (81k Answer/192k Reasoning) in 5.2.
There's a whole frontier of cost vs performance and this model pushes it out in a bunch of places, so its not just a matter of making a different tradeoff as you make it sound.
This model is on the cost performance frontier (Pareto line) from medium -> max.
If you leave xhigh and max off.. you still get better performance than Opus 5.5 max
Opus 5.5 Max - score 59 - cost per task - $10.04
Fable 5.2 High - score 62 - cost per task - $9.08
Note: This doesn't really capture the reality of things like 3d design or electronics design, where the earlier models just can't DO any of it well. Maybe you can bandaid over some of the gaps with strong skills, etc. The newer models can just do things that were not possible at all. When you move outside of coding into the visual/spacial space, you'll feel the inflection point on those capabilities that are happening right now.
The thing is max reasoning is most effective at things which are least interesting (the ones where the answers can be validated by a computer.) I've found for a lot of tasks like image recognition, low reasoning can even be better quality than max.
- Sol-low - cost-competitive with Haiku, while outperforming Sonnet on medium
- Sol-High - Competes with Astra medium and beats Opus 5.5-medium, at 20% the cost
- Sol-Max - Beats Opus 5.5 high, Astra high, and competes with Opus xhigh, at cheaper than Sonnet and a fraction of the Opus cost
And when you look at time spent, With the exception of Sol-max you get better results in less time for less money. My own experience is that this is true-- it is difficult to call Opus strictly better than Sol, but Sol gets where it's going far more efficiently.
It's rather hard to look at opus-- or any of anthropic's models-- and say "economical".
On the other hand, there are plenty of ethical reasons not to support Sam Altman/Peter Thiel/Larry Ellison in their particular world vision, which for many rules out use of OpenAI.
They have not run into a wall. The frontier labs distill their strongest models into their cheaper, faster models.
Having these highly rated models reflects (potentially) the frontier of what they are capable of producing. Essentially its and advertisement and shows directionally where the smaller faster model's capabilities are headed.
Maybe a dumb question but once you have a model that spends a lot of tokens for high performance why can’t you distill it into a model that thinks less?
Sure it wouldn’t generalize well to new tasks but plenty of tasks are important and recurrent enough to justify this.
There's also a bunch of RL research around strategies to coax the student model towards more concise trajectories.
In general, a lot of people in industry are finding cost savings these days by tuning a smaller model over longer agent trajectories (not exactly the same as distilling the reasoning traces, since those aren't exposed by many frontier labs anymore) collected from their production agents. Typically works best when you have a pretty workflow-y agent whose tasks all share a similar shape.
Caleb is right.
Also it seems the results are not straightforward.
Let say you have a model that score 100 in a benchmark and you distill it into a model half the size. It has very few guarantees that it will score close to 100 again.1
But I assume one of the hard problems is to know that the model might not perform well enough outside of the distillation part.
There is also another caveat that models do display some level of sub language control, by this I mean that innocuous text might contain information relevant for the own agent or for other agent with it shares tasks.
(Like when your partner says there is nothing wrong)
Because distillation uses very few resources, we don't know exactly what gets send forward or not.2
As Caleb said, many companies found that they can tune small open weight agents over procedural known tasks because that's easy to inspect and check and because there is little space dor ai-loitering, it tends to be less problematic.
The key difference, at least in my understanding, is that you are not concerned with the weights or reducing data. You are basically tuning agents existing weights to emulate another agent answer.
1 - The process of distillation is super effective for specific domains though.
2 - I know I will be called out for supposed AI mystery machine. What I mean is that Large models rely on latent reasoning and nuanced context, often lost during condensation. Distilling models to use fewer tokens or parameters makes predicting retained details difficult.
Some papers suggest that model preferences introduced during training, whether intentionally or inadvertently, may be inherited by distilled and fine-tuned models.
We all knew it was coming. There’s only so much space with current capacity, speeds, and hardware.
We would need to change how we compute. New chips. New architectures. Or maybe we reached the end of physics. Photonic compute is the last frontier. I don’t know.
I don't think the problem is computing.
The method has always had a fatal flaw.
Even if they double computing and time, they will not have enough content and even if they did, probably they would not increase proportionally.
They are doing IPO now because it might become apparent that they reached a limit in the next year
But if the intelligence has been calculated once it can then be utilized for the future. It might still make sense for cases in which you are looking for breakthroughs. I would assume the intelligence is gonna trickle down.
Diminishing returns would catch you there as well.
Perhaps, it's a bit like being a super smart person in contemporary society: at some point, being smarter doesn't really reliably translate into a better life.
When you learn from available human data, those text books max out at some point. You cannot learn what isn't there.
I don't necessarily think this is true though. The hype cycle, and the false promises are what most people here dislike it seems, along with the snake oil salesmen telling us "this time it really is different!"
And I am simply ignoring the ridiculous comments around local model usage from certain CEOs.
That would be my layman's assumption as well. A theory is they're just releasing this as a stepping stone towards something bigger before the IPO. Either way it's the same way chinese models get higher scores on benchmarks. Drop enough of wait actuallys and you'll be top the charts, but time per task and tokens per task go up by a mile (as is the case with this release).
How do you explain Opus 5.5 or GPT 6.1? I could maybe agree if this is a general trend. But one "supermodel" existing that is really expensive does not signal diminishing returns.
If you compare Fable 5.2 Medium to Opus 5.5 high (settings that make sense - max is universally considered rarely worth it), you will see that Fable 5.2 is SUPPOSEDLY very cost competitive: https://artificialanalysis.ai/models/claude-fable-5-2?models...
The charts only show the cost per task - not whether or not they got the task correct - which makes flash models look way better than they actually are.
Yes, but that's running at max. Nobody is going to do that unless they want the best possible performance at all cost, in which case, they get what they asked for.
Fable 5.2 at high scores a 62, which is better than anything else on the market (except Fable 5.2 at xhigh and max) -- and is about a third of the price per task.
The model feels bench maxed to me. It is back to Opus and prior Fable behavior of me pointing out constant mistakes, with it apologizing and saying that it should have looked at what I point it to.
The thing is, all of these pointers are in the CLAUDE.md file, so it just ignored instructions because it thought it was smarter than the world and didn't need them. It feels really arrogant is how I would describe it.
Its still prohibitively expensive, and models 5% the cost are faster and get the job done just as well (for the average person's workflow).
I really don't get the value proposition of these massive models, seems like they're just loss leaders simply to say you're the in the top 3. There's not much use cases outside of niche academic research (and even then I question the need). They mostly function as marketing tools, like a casino that offers high roller accommodations.
People will say "cyber" but companies didnt want to spend on security before, not sure why they would all of a sudden be compelled to spend even more.
Not everyone is writing CRUD business apps and personal use software. Pretty much any model can do those now. The frontier is for doing things those models struggle with. Even in vibe coding a simple video game the difference in results between an open source model, the previous frontier model, and the current model is pretty huge.
Security is mandatory now in a way that it wasn't, now that AI has eliminated the skill barrier to hacking.
I think the general notion that a lot of people are using models that are overpowered for their use case is right, but I think the big question is where the line actually is and how many people fit on either side. My suspicion based on my experience is that there are significantly more people on the "actually need the frontier model" side than the bear case suggests, but I'm not speaking with any authoritative data here and could be totally wrong.
I used to work on evals/observability at Opik: https://github.com/comet-ml/opik And I can say from working with lots of users running agents at scale, there is definitely a pretty commonly reached frontier of difficulty/complexity/abstraction where the capability difference between something like Fable/Astra and smaller models is a big deal. I find this to be especially true the more autonomous your workflow gets. You can still often tackle the hardest tasks with something like Opus, it just typically involves more direct intervention and design work from the user.
But again, this is just anecdata on my part. All of this is still pretty ambiguous and lots of companies are trying to solve the problem of matching models to task difficulty (I.e. routers), with varying degrees of success.
People looking at frontier models release are not the average person, and let's say that you pay $150 with Fable to save $1000*, it's much more interesting than using $10 with a cheap model to save even $500.
And will also be drinking coffee, and using a computer, text editor, browser, search engine, etc., surely we don't start attributing accomplishments to the tools used?
There have been many incredible breakthroughs in cancer in recent years without it. If the AI money had gone into that, I suspect we'd have better results.
AI seems more likely to help with detecting cancer than curing it imo.
I’d liken the expensive models to the early mainframes, or the first colour TV cameras. Sometimes it’s worth paying a fortune for the best thing available even if (compared with what’s to come) it’s not very good yet.
Out of interest, do you run the same tests more than once to generate that?
What I often observe is that models being asked the exact same prompt can alternate between getting something right and getting it wrong, thanks to the wonders of a non-deterministic output.
I think we will beat the cost with Hardware. Faster, more efficient hardware. From what I have seen we have only found ways to make it smarter and not more cost effective.
I'm curious to see the next generation of dedicated AI processors and how they perform.
Unfortunately, I came to the conclusion that higher intelligence is extremely beneficial even for 'dumb tasks'. So I resorted to switching to Opus for pretty much everything.
to me , glm flash and deep seek is that chinese 3 wheel tuk-tuk or e-bike runs on electricity. It make perfect sense to run to supermarket, and it's a backbone of Chinese economy, but no way in my life for a chance to save $100/mo you will risk your business with bugs, and bad/worse architecture decisions to save that $100.
again, they can be great for narrow use cases, and save you a fortune.
PRC models are more like ubiquitous class8 trucks that all economies can use as backbone. SOTA AI is like sending a courier on a first class ticket, only good for niche high ROI scenarios. No one is haggling over $100 per month, but $1000s per head per month on enterprise SOTA vs PRC models. For subsidized $100 per month plans SOTA vendors goes broke anyways... at premium SOTA API most users go broke. Pareto frontier PRC models are already being routed to take over all the general use cases, seems like essentially everyone outside of niche coding is scaling back SOTA AI use because it's not worth the $$$. It didn't take long at all for people to figure out you don't need next day airfreight everywhere.
The frontier chart there makes gemini carbon look really weird. Low is worse and more expensive than medium, and high is worse and more expensive than medium... funky curve like that implies a messed up model or weird benchmark.
It beats opus 5.5's max effort with less cost, which is a step forward if you don't want to crank it to Ultra Max Settings.
The amount of compute it cost to get the best previous performance from anthropic went down. Nobody is putting a gun to your head and forcing you to go further down the paretto frontier if the current level of intelligence/task is acceptable to you.
Those are both true, but I don't see how they are meant to rebut what I said.
Most people on plan use do NOT use up all their plan usage.
Users on API plans can still use Opus 5.5 which is a great model. If they want top of the line performance, they can pay more.
When another company puts out a model that undercuts Fable 5.2 on price at similar performance, Anthropic will put out Opus 6 and undercut them right back. That's how the wheel turns. This is just part of the cycle.
Most is doing the heavy lifting here. I am a power user. And I max 3 claude max subscriptions on opus 5.5 high and a single codex pro 5x for personal use.
This has echoes of the same argument that why hire anybody in the US when you can hire someone almost as good elsewhere for a small fraction of the cost.
In winner-takes-all markets, it becomes worth it. The utility of the marginal improvement doesn't actually diminish, in fact, it increases like a hockey stick. If you could pay 3x more in compensation to hire a 20% better employee, it's almost always worth it in knowledge work industries.
I don't marginal intelligence has diminishing value -- quite the opposite actually. Every additional "IQ point" or 9-of-reliability unlocks some new economically valuable task. But in a few years we'll know whether or not that's correct!
I am always interested in this comparison since it shows how quickly you can go back and forth with a model, capturing both tokens/task and tokens/second. The new models seem to be meaningfully better here.
Note that the graph is misleading for open models since it's only measuring their official providers and that's rarely the fastest.
For example, the official provider for Kimi K3.1 Pro is dead last for speed on OpenRouter. You could run it more than 5x faster for the same cost elsewhere, which should push it comfortably into the pareto quadrant.
Why does AA choose ~$3.1 per task to be the breaking point for "most attractive quadrant" highlight? Surely the most attractive section is just being close to the entire frontier?
The choice of X-axis maximum and Y-axis maximum dictates the cutoff point of the quadrant. With a maximum of $80 per task, the quadrant is $3.1 and lower. If some new model gets released and costs more than $80 per task, the "most attractive" cutoff moves up from $3.1 since they have to move the X-axis max. It's completely arbitrary.
It is suggesting a hard cutoff since it says "most" attractive, a superlative, instead of "more" attractive. And it is not pareto optimal, because if the models get really good and are scoring 10x more for 10x cheaper that green region will no longer be pareto optimal.
Consider why Hollywood movies use $100k lenses, rather than $500 lenses. Are the lenses really 200x better? Probably not, but there are cases where that extra 5% better is worth an extra $99.5k. There will be things Fable 5.2 can do reliably that Opus 5.5 can't; and there will be people willing to pay for that.
I think those öensrs are mostly rented. For a $100M film few thousand a day is nothing. Those lenses also come with a camera. If you want to use same camera and lens in a desert in Tunis and snow storm in Norway and get same quality of image without constantly changing broken equipment, you are going to either pay thousands for a good equipment or millions for delayed timetables for broken equipment. In that sense, it is not only 200 times better, it is your only option. So yes, sometimes the price is worth it.
Furthermore, you make one lense and rent it out, apparently some Panavision lenses made in 60’s are still in use. It doesn’t churn constantly electricity and GPU-chips sucking in more and more resources like these things do.and at the same time, it is quite hard to find the exact cases where the markup is justified, unlike with those lenses. Maybe in cybersecurity etc. but even there some Chinese model can be almost equally good or even better.
Those lenses save money in budgets and make money in getting people into theatre or streaming services to watch films creating profits way beyond their cost. I think in terms of economics that is productivity increase justifying the cost. We are still waiting for that productivity increase come from the AI industry.
Modern studio pictures that get nominated for best picture or best cinematography cost around $80 million on average. If you look at highest grossing films the budgets are closer to $175 million on average – even taking into account the $20 million budget of Demon Slayer: Infinity Castle.
Isn't it more that when you're spending that kind of money that $100k for the "best" thing becomes kind of like "insurance" where you don't want to be the one who doesn't use "the best" when "the best" costs about 0.1% of the picture's budget?
Additionally a studio can buy the $100,000 lens once and use it on many films. But it's not like you can reuse Fable 5.2 in the same way.
Almost no variant of the same product that costs 100x than another is actually 100x better than the other variant. Whether the increase in cost is worth the increase in value depends on the application/use case. Most often the higher priced tier is not worth it for most use users, but most often there are still some use cases where it's desired. This applies to almost anything you can buy: food, wine, clocks, cars, planes, bikes, ski, clothes, furniture, phones, appartments, insurance, service, anything. Surely also LLMs.
the concept of diminishing returns exists in every thing. headphones, coffee, cars etc. anything that can be pretty continuous in terms of quality will have this
You too can sell stuff for $100k. Just make something useful to someone who can throw around $100k.
You think a $100K lens is crazy? Wait until you hear what people are paying for enterprise software. Big business will spend 10x that on software... and then never implement it.. The lens is a deal.
It is about time to start setting taxes for ai usage in commercial settings and set how to transition to a society with less available work. They talk so much we need universal benefits, about time for them to start paying for it
https://artificialanalysis.ai/#intelligence-efficiency-tabs this is the chart that matters for most of poor developers and companies. Pareto for a given intelligence index and looks like only openai is across the quadrant with muse at the bottom. They have really cooked efficient models!
Idk how could Fable 5.2 win the benchmarks. In my tonight trials it drifted quite a lot. I was analyzing the feature's design with it, and it lied to me several times, ending up with "You're right." Opus 5.5 was better.
Later, I ran a "health check" session—asking "what went wrong". I read the session, and its suggestion turned out to be something that was already part of the design.
It’s like the attention mechanism is too leaky or something.
I'm finding it completely stupid compared to Opus 5.5. It doesn't seem to be following required rules in my setup which is for non-coding purposes. It basically made up results when I have a whole system with papers that are knowledge rag searchable and it is required to search there, but just decided not to.
Thanks, I may end up giving them a try. I actually use Codex extensively for the custom plugin that tells Claude what needs to happen with hooks for reminders, some commands require a search before they do a grep (which Claude seems to love to start with), etc.
Long execution time makes this not very impressive. It's easy to benchmaxx if you give the model an insane compute budget and room to just brute force the problem. Efficiency is what actually matters.
I am wondering which use cases would require usage from Fable 5.2 especially when Opus 5.5 already performs almost all of the tasks in a very optimized way.
Ask opus to write a occlusion mapper shader and create a 3d environment where ground texture is used to mimic infite detail. Same on some random brick walls. Corners need to be extruded realistically.
This is a hard task, because only a few games have this (crimson Desert is one of them)
Fable, for me (it might be a completely different experience for you) has been like the well read snobbish kid at school that openly ignores me when I ask certain questions and often asks me to talk to someone else on other questions.
That kid might indeed be a genius, clearly bright, but is more interested in deciding whether I should have asked a question than in answering it.
For me personally, if the best model on the planet refuses to answer why I shouldn't mix bleach and ammonia, where I'm trying to understand the chemistry behind the warning, it's not pareto frontier to one that atleast tries to and points me in a general direction. Being told that the topic is concerning, referred to someone else and charged part of my quota for that experience feels like getting an invoice from that school genius for telling me to "just ask the chemistry teacher".
If my LLM asks me to rethink my life when I ask it to reverse engineer the firmware of my Samsung TV that I suspect is setting up adhoc radio networks to relay my data to Samsung's cloud, I feel rather sad, specially when that kid sends me an invoice for saying "no"
At least Anthropic should not bill me for the refusal (as of my last tests, refusals still comes out of my quota) - getting slapped and pickpocketed for asking a question is really upsetting.
fable is a high agency model (some people call it 'big model smell'). if you want something that blindly follows instructions, use sonnet. I want to be challenged by the thing since I expect it to be smarter than me at everything except the little part of human knowledge I'm an actual expert in.
not really, no - but I almost exclusively use the biggest models as advisor and reviewer subagents, so if they refuse to their smaller siblings I don't get to see it
Similar experience with Opus. Asked it to tighten access controls and isolation in my homelab. Designed something with a whole bunch of words that looked plausible, so I told it to go ahead. Halfway through it gets stuck because it locked itself out of something, and asks me to run some commands as root.
Me: No, I'm away from the computer right now, you figure it out. Use the root login you already have if you can't figure it out.
Opus: This can wait until you get home, then.
Me: No, figure it out now or use root.
Got [cyber]'d, some scary message that referred me to legal/AUP, and refused to continue.
Switched to OpenAI in part due to this silliness. Sol seems a little less smart but doesn't get on a high horse to lecture you about this stuff. It either does the work or fails to do the work, and leaves the moralizing out of it. It was able to take over from where Claude left off and finish the project.
Google could have been #1 for a little while, but they botched their release and squandered their temporary lead. (As usual.) If I worked there, I'd be so pissed.
Yes, I have found Argon is good at explaining things, which is nice in root cause analysis and debugging / troubleshooting
For an open source project (CrowdGPT) I hit upon this workflow in a recent PR:
DeepSeek v4 Flash 1017 made a draft plan
Google Gemini 4 Argon corrected the plan
DeepSeek v4 Flash 1017 did the first implementation
Google Gemini 4 Argon debugged the problems in that implementation
Kimi 2.7 Code did a code review and proposed minor fixes
DeepSeek v4 Flash 1017 performed the fixes
Google Gemini 4 Argon debugged one problem that I found in that fix.
I reviewed the plans and the resulting code
Pretty big Gemini fan - but IMO Argon is universally worse for coding than Astra or Opus 5.5 (un-nerfed).
I strongly prefer Gemini's output compared to anything else. So I typically use Argon to either direct other agents or explain something without psychobabble.
Argon/Carbon have by far the lowest hallucination rates, so IMO it serves as the best bullshit shield from other models.
In terms of implementation (and especially design), it's far behind Astra and Opus, and I'm sure Fable will make that gap even larger.
Argon/Carbon may give you not-the-best advice, but they are far less likely to just flat-out-and-completely lie to you.
I do. I prefer it to Opus 5.5 - just as good at coding for my needs, but so much more pleasant to work with.
I am not sure how people keep saying google are not in the race, or have lost it or whatever when just a few hours ago they had the #1 model? What more do people want?
I actually just think it's a meme to talk down on Gemini at any opportunity here. There's probably a fair number of people that wrote off antigravity at its (pretty rocky) launch. And Google also doesn't provide any kind of good sandbox-y experience to Just Try their models, which probably bites them here too.
But Gemini models are still neck and neck in the lead both in my experience and in benchmarks, and Gemini is far above the competitors (although Opus 5.5 closed much of the gap) when it comes to any kind of complex web dev. (I generally use Astra for anything 3D, Gemini for anything web/mobile, and Opus for everything else.)
Their latest internal reshuffling and the strange amount of parallel releases with the same target, their kafkaesque barriers to entry, and a lack of clear vision is what makes people think they’re not in the race anymore. Obviously they still are. They develop new models and also are deep in the compute game too.
In my experience Google models have always been mediocre at best and dangerous to use at worst but I must admit that I didn’t try the latest Aragon/carbon crop.
I learned not to put too much trust into benchmarks. The best benchmark is to use the model on a codebase you know well.
Gemini 4 Argon is available in both anti-gravity and the Gemini app for me on my Gmail account and I didn't do anything special (I have a modest drive storage subscription - not the largest nor smallest... Not sure if it is bundled with that or not).
So certainly released to the unwashed masses like me to use for free without any special extra work or API keys.
I tried using it as part of my Pro/One/whatever subscription (the one that comes with Youtube Lite). For code review, it was able to find some things that Sol 6.1 couldn't. I found it quite useful. But for code generation, it's pathetically slow and the code isn't impressive. It's at least one year behind.
Optics and incentives. Google will turn on the R&D the day after their Anthropic investment is realised, maybe $80-150bn on a 1.2tn valuation, plus >100bn in spending commitments.
Reminds me a lot of the race for clock speed when I was a kid. We just took it for granted that it was going to keep increasing, but CPUs sat at roughly 1-2GHz for a long time, or at least that's how it seemed. The slowdown and pause was fairly noticeable. The current release cadence is unsustainable. It might be that we're getting close to the AI equivalent of 1-2GHz CPUs, and it takes years of incremental progress to push further.
When AMD won the race to 1 GHz, the GHz race slowed and Intel and AMD began focusing more on multi-core. Some chips continued to push the boundaries up to several GHz, some quite higher with specialty cooling. But multi-core was the far better strategy.
There's only one problem: what good is it if you can never use it? What I mean by that is, using it reliably is not possible on a, day-in-and-day-out basis to build on Anthropic and then expect it to be stable over time.
It's insane how people nitpick on the highest thinking setting for every anthropic release and claim it's too expensive. Not trying to stan for anthropic, but what am I missing here: Just don't use those models on the highest thinking setting?
Fable 5.2 medium for example seems a really good choice when it comes to the performance vs task (if you have a complex task that requires such a model of course).
This has got to be a minor reason why adoption of AI isn’t faster.
Businesses like predictable costs and predictable performance.
If your new model is slightly better but crazy expensive, but doesn’t have to be with certain settings, but my user can accidentally run it that way by mistake or maybe your harness makes it the default by accident and the user doesn’t notice…
My beloved predictability is destroyed and trying to get users to understand this is useless.
Imagine a streaming subscription that worked like this: “We’ve updated our catalog with more movies. If you want the premium ones though you only get 2 per week and then you have to wait. But if you do mostly classics and a few from the 2010s you can squeeze in a recent theater release. If you time it right and do only one every day you can see 30 modern classics this month, but you can’t watch three today. Also your bill this month is $40 instead of $20 due to the library update”.
Yep, and now multiply that complexity by hundreds or thousands of employees. That's why companies are giving in to crude, hacky solutions like monthly billing caps. What else are you really gonna do?
The output tokens per task just exploded. From 118k tokens (46k Answer/71k Reasoning) in 5.1 to 274k tokens (81k Answer/192k Reasoning) in 5.2.
Notably, in order to achieve a specific quality, the model has to "reason" for every prompt proportional to the training inference.
This is arguably the case of diminishing returns being reached, as answer token prices saturate the willingness of customers to pay.
They've run into a wall.
An economically successful step would be to improve training cost/frontier advance.
They essentially failed at that, assuming they kept their training costs for the model step at least constant.
If you leave xhigh and max off.. you still get better performance than Opus 5.5 max
Opus 5.5 Max - score 59 - cost per task - $10.04
Fable 5.2 High - score 62 - cost per task - $9.08
Note: This doesn't really capture the reality of things like 3d design or electronics design, where the earlier models just can't DO any of it well. Maybe you can bandaid over some of the gaps with strong skills, etc. The newer models can just do things that were not possible at all. When you move outside of coding into the visual/spacial space, you'll feel the inflection point on those capabilities that are happening right now.
Notable takeaways:
- Sol-low - cost-competitive with Haiku, while outperforming Sonnet on medium
- Sol-High - Competes with Astra medium and beats Opus 5.5-medium, at 20% the cost
- Sol-Max - Beats Opus 5.5 high, Astra high, and competes with Opus xhigh, at cheaper than Sonnet and a fraction of the Opus cost
And when you look at time spent, With the exception of Sol-max you get better results in less time for less money. My own experience is that this is true-- it is difficult to call Opus strictly better than Sol, but Sol gets where it's going far more efficiently.
It's rather hard to look at opus-- or any of anthropic's models-- and say "economical".
Having these highly rated models reflects (potentially) the frontier of what they are capable of producing. Essentially its and advertisement and shows directionally where the smaller faster model's capabilities are headed.
Doesn’t seem like a “wall” to me.
This is a dataset basically constructed by taking a bunch of DeepSeek R1 trajectories and filtering out the incredibly long rambling ones: https://huggingface.co/datasets/open-thoughts/OpenThoughts-1...
There's also a bunch of RL research around strategies to coax the student model towards more concise trajectories.
In general, a lot of people in industry are finding cost savings these days by tuning a smaller model over longer agent trajectories (not exactly the same as distilling the reasoning traces, since those aren't exposed by many frontier labs anymore) collected from their production agents. Typically works best when you have a pretty workflow-y agent whose tasks all share a similar shape.
Let say you have a model that score 100 in a benchmark and you distill it into a model half the size. It has very few guarantees that it will score close to 100 again.1
But I assume one of the hard problems is to know that the model might not perform well enough outside of the distillation part.
There is also another caveat that models do display some level of sub language control, by this I mean that innocuous text might contain information relevant for the own agent or for other agent with it shares tasks. (Like when your partner says there is nothing wrong) Because distillation uses very few resources, we don't know exactly what gets send forward or not.2
As Caleb said, many companies found that they can tune small open weight agents over procedural known tasks because that's easy to inspect and check and because there is little space dor ai-loitering, it tends to be less problematic. The key difference, at least in my understanding, is that you are not concerned with the weights or reducing data. You are basically tuning agents existing weights to emulate another agent answer.
1 - The process of distillation is super effective for specific domains though.
2 - I know I will be called out for supposed AI mystery machine. What I mean is that Large models rely on latent reasoning and nuanced context, often lost during condensation. Distilling models to use fewer tokens or parameters makes predicting retained details difficult. Some papers suggest that model preferences introduced during training, whether intentionally or inadvertently, may be inherited by distilled and fine-tuned models.
To do that you need a larger model to have many many branches to try.
We would need to change how we compute. New chips. New architectures. Or maybe we reached the end of physics. Photonic compute is the last frontier. I don’t know.
They are doing IPO now because it might become apparent that they reached a limit in the next year
And maybe that's how all of it is going to evolve in the future – so you pay per solution, not tokens?
Perhaps, it's a bit like being a super smart person in contemporary society: at some point, being smarter doesn't really reliably translate into a better life.
When you learn from available human data, those text books max out at some point. You cannot learn what isn't there.
And I am simply ignoring the ridiculous comments around local model usage from certain CEOs.
Fable 5.2 low is 3x cheaper than Fable 5.1 xhigh while being 8% smarter.
https://artificialanalysis.ai/?releases=claude-fable-5-2%2Cc...
Fable 5.2 at high scores a 62, which is better than anything else on the market (except Fable 5.2 at xhigh and max) -- and is about a third of the price per task.
The thing is, all of these pointers are in the CLAUDE.md file, so it just ignored instructions because it thought it was smarter than the world and didn't need them. It feels really arrogant is how I would describe it.
I really don't get the value proposition of these massive models, seems like they're just loss leaders simply to say you're the in the top 3. There's not much use cases outside of niche academic research (and even then I question the need). They mostly function as marketing tools, like a casino that offers high roller accommodations.
People will say "cyber" but companies didnt want to spend on security before, not sure why they would all of a sudden be compelled to spend even more.
Security is mandatory now in a way that it wasn't, now that AI has eliminated the skill barrier to hacking.
I used to work on evals/observability at Opik: https://github.com/comet-ml/opik And I can say from working with lots of users running agents at scale, there is definitely a pretty commonly reached frontier of difficulty/complexity/abstraction where the capability difference between something like Fable/Astra and smaller models is a big deal. I find this to be especially true the more autonomous your workflow gets. You can still often tackle the hardest tasks with something like Opus, it just typically involves more direct intervention and design work from the user.
But again, this is just anecdata on my part. All of this is still pretty ambiguous and lots of companies are trying to solve the problem of matching models to task difficulty (I.e. routers), with varying degrees of success.
*Scale numbers to whatever suits you.
That's about all you can say for sure, especially about a topic that is so broad as to be virtually meaningless.
Really?
AI seems more likely to help with detecting cancer than curing it imo.
There have also been many incredible breakthroughs in math in recent years without AI. But things are different now.
If the AI money had gone into that, I suspect we'd have better results.
Otherwise known as "eating your seed corn."
Compared to other new technology, like, say mRNA vaccines.
1 - https://bench.killswitch-lang.org/
also, very cool idea
What I often observe is that models being asked the exact same prompt can alternate between getting something right and getting it wrong, thanks to the wonders of a non-deterministic output.
I'm curious to see the next generation of dedicated AI processors and how they perform.
Anthropic and OpenAI are expensive bleeding edge, but few use cases really need it.
again, they can be great for narrow use cases, and save you a fortune.
The amount of compute it cost to get the best previous performance from anthropic went down. Nobody is putting a gun to your head and forcing you to go further down the paretto frontier if the current level of intelligence/task is acceptable to you.
They've been using Fable as a selling point, but upping to this model will presumably eat further into their profits (should any exist).
Cheaper and faster than Opus(max) AND scored higher. But yeah let's doompost about it...
Booooooooooo!
Most people on plan use do NOT use up all their plan usage.
Users on API plans can still use Opus 5.5 which is a great model. If they want top of the line performance, they can pay more.
When another company puts out a model that undercuts Fable 5.2 on price at similar performance, Anthropic will put out Opus 6 and undercut them right back. That's how the wheel turns. This is just part of the cycle.
This might be true for a small portion of the user base, but it is not true as a statement of fact.
https://support.claude.com/en/articles/9266767-what-is-the-t...
> Team plans support up to 150 seats. If your organization is on the Team plan and needs more than 150 seats, you can upgrade to the Enterprise plan.
https://claude.com/pricing
> Enterprise: Seat price + usage at API rates
> US$20/seat/month, billed annually, 20 seat minimum. Usage cost scales with model and task.
There is no premium seat option listed for the enterprise plan
While they made, hands down, the best models, their 100x price for few percent better is completely made up, and I think it will end up bad for them
In winner-takes-all markets, it becomes worth it. The utility of the marginal improvement doesn't actually diminish, in fact, it increases like a hockey stick. If you could pay 3x more in compensation to hire a 20% better employee, it's almost always worth it in knowledge work industries.
I am always interested in this comparison since it shows how quickly you can go back and forth with a model, capturing both tokens/task and tokens/second. The new models seem to be meaningfully better here.
For example, the official provider for Kimi K3.1 Pro is dead last for speed on OpenRouter. You could run it more than 5x faster for the same cost elsewhere, which should push it comfortably into the pareto quadrant.
If you take the average point inside the 4 quadrants, the top left one is pareto optimal (of the 4)
Another member of the pro-REPL / anti-agentic tribe. There are dozens of us!
But perhaps the performance on real-world tasks will be more noticeable than on benchmarks.
Furthermore, you make one lense and rent it out, apparently some Panavision lenses made in 60’s are still in use. It doesn’t churn constantly electricity and GPU-chips sucking in more and more resources like these things do.and at the same time, it is quite hard to find the exact cases where the markup is justified, unlike with those lenses. Maybe in cybersecurity etc. but even there some Chinese model can be almost equally good or even better.
Those lenses save money in budgets and make money in getting people into theatre or streaming services to watch films creating profits way beyond their cost. I think in terms of economics that is productivity increase justifying the cost. We are still waiting for that productivity increase come from the AI industry.
Modern studio pictures that get nominated for best picture or best cinematography cost around $80 million on average. If you look at highest grossing films the budgets are closer to $175 million on average – even taking into account the $20 million budget of Demon Slayer: Infinity Castle.
Isn't it more that when you're spending that kind of money that $100k for the "best" thing becomes kind of like "insurance" where you don't want to be the one who doesn't use "the best" when "the best" costs about 0.1% of the picture's budget?
Additionally a studio can buy the $100,000 lens once and use it on many films. But it's not like you can reuse Fable 5.2 in the same way.
You too can sell stuff for $100k. Just make something useful to someone who can throw around $100k.
You think a $100K lens is crazy? Wait until you hear what people are paying for enterprise software. Big business will spend 10x that on software... and then never implement it.. The lens is a deal.
Fable (5.1 and 5.2) consistently write a much shorter script (2.5 pages) compared to Opus 5.5 (4.5 pages).
Fable 5.2: https://claude.ai/share/274a96a8-c8a5-4564-9ca1-e914edffc722
Opus 5.5: https://claude.ai/share/da8ab6bd-0987-47dc-9c26-68ebbfb25781
Edit: high seems better than max
Later, I ran a "health check" session—asking "what went wrong". I read the session, and its suggestion turned out to be something that was already part of the design.
It’s like the attention mechanism is too leaky or something.
I remember Mythos 5.1 being a tad better even for not cybersecurity tasks.
That kid might indeed be a genius, clearly bright, but is more interested in deciding whether I should have asked a question than in answering it.
For me personally, if the best model on the planet refuses to answer why I shouldn't mix bleach and ammonia, where I'm trying to understand the chemistry behind the warning, it's not pareto frontier to one that atleast tries to and points me in a general direction. Being told that the topic is concerning, referred to someone else and charged part of my quota for that experience feels like getting an invoice from that school genius for telling me to "just ask the chemistry teacher".
If my LLM asks me to rethink my life when I ask it to reverse engineer the firmware of my Samsung TV that I suspect is setting up adhoc radio networks to relay my data to Samsung's cloud, I feel rather sad, specially when that kid sends me an invoice for saying "no"
At least Anthropic should not bill me for the refusal (as of my last tests, refusals still comes out of my quota) - getting slapped and pickpocketed for asking a question is really upsetting.
There's a setting to not degrade to a lower model on refusals and instead pause the chat.
Me: No, I'm away from the computer right now, you figure it out. Use the root login you already have if you can't figure it out.
Opus: This can wait until you get home, then.
Me: No, figure it out now or use root.
Got [cyber]'d, some scary message that referred me to legal/AUP, and refused to continue.
Switched to OpenAI in part due to this silliness. Sol seems a little less smart but doesn't get on a high horse to lecture you about this stuff. It either does the work or fails to do the work, and leaves the moralizing out of it. It was able to take over from where Claude left off and finish the project.
Max effort is nice for revealing the limits of a model, but it's almost never used in practice.
If you scroll down to the coding charts you can start adding Fable 5.2, Opus 5.5, and others at more normal difficulty settings.
Fable 5.2 on medium looks quite attractive compared to Opus 5.5 high/xhigh
Fable 5.2 on max (the default) looks like a good way to light your money on fire.
For an open source project (CrowdGPT) I hit upon this workflow in a recent PR:
Pretty big Gemini fan - but IMO Argon is universally worse for coding than Astra or Opus 5.5 (un-nerfed).
I strongly prefer Gemini's output compared to anything else. So I typically use Argon to either direct other agents or explain something without psychobabble.
Argon/Carbon have by far the lowest hallucination rates, so IMO it serves as the best bullshit shield from other models.
In terms of implementation (and especially design), it's far behind Astra and Opus, and I'm sure Fable will make that gap even larger.
Argon/Carbon may give you not-the-best advice, but they are far less likely to just flat-out-and-completely lie to you.
I am not sure how people keep saying google are not in the race, or have lost it or whatever when just a few hours ago they had the #1 model? What more do people want?
But Gemini models are still neck and neck in the lead both in my experience and in benchmarks, and Gemini is far above the competitors (although Opus 5.5 closed much of the gap) when it comes to any kind of complex web dev. (I generally use Astra for anything 3D, Gemini for anything web/mobile, and Opus for everything else.)
I learned not to put too much trust into benchmarks. The best benchmark is to use the model on a codebase you know well.
So, what more do I want? It'd be a good start to actually tell me it's available and giving me a clear path to using it via plan & harness?
well, releasing it would be nice
So certainly released to the unwashed masses like me to use for free without any special extra work or API keys.
It's not as smart for coding as Opus, but it's above the best open weights.
IME, YMMV, etc.
I had to remind them 4.6 only came out February this year. In a normal world 4.6 is a newer model! A model 6 months old is seen as ancient!
In the JS ecosystem 6 months always meant breaking changes, without incremental updates everything would be basically like a full port
Pretending like I did “20% hands on coding” just doesn’t cut it, even the reality is having no idea what’s going on anymore
The crypto space moved that fast as well
And AI keeps the same cadence
It’s more important for me to be able to build
Max should not be used unless you really need every bit of perf, it's deep into diminishing returns territory.
Change the models in the dropdown to medium / high and Fable 5.2 is about twice as cost efficient, and slightly faster.
Fable 5.2 medium for example seems a really good choice when it comes to the performance vs task (if you have a complex task that requires such a model of course).
Businesses like predictable costs and predictable performance.
If your new model is slightly better but crazy expensive, but doesn’t have to be with certain settings, but my user can accidentally run it that way by mistake or maybe your harness makes it the default by accident and the user doesn’t notice…
My beloved predictability is destroyed and trying to get users to understand this is useless.
Imagine a streaming subscription that worked like this: “We’ve updated our catalog with more movies. If you want the premium ones though you only get 2 per week and then you have to wait. But if you do mostly classics and a few from the 2010s you can squeeze in a recent theater release. If you time it right and do only one every day you can see 30 modern classics this month, but you can’t watch three today. Also your bill this month is $40 instead of $20 due to the library update”.
Nag, fuggit. Let's just burn all the capital in the world and ride it into the flames.