They have data from their competitors model outputs. It is very hard to serve an LLM without also exposing how it works.
Combustion engines improved gradually, each year. One year they got better than horses.
The assumed timeline (2 months) is slightly wrong because Fable (Latin) is essentially the same as Mythos (Greek) albeit with protections against cyber and biological misuse.
Mythos (Preview) was publicly announced in April 2026 [1] which means other labs have had 4 months to catch up, not 2 months.
Assuming everyone had access to Mythos from the start, your expression, similar to other folks would have been "Mythos-level intelligence" and not "Fable-level intelligence".
Well that’s not really true; it covers completely legitimate use also.
Note that RLVR is incredibly compute expensive but it's CPU as much as GPU.
I realise this might be a skill issue.
I prefer models that are less "smart" but faster. Do the thing I asked you to do, immediately, and if you can't tell me and we'll work it through. Iterate faster not smarter.
I'm not stating this as a fact, but it's a hypothesis I'm keeping in my mix.
Other labs catching up in half a year seems about right.
It's probably a mix of all of that plus simply always keeping one in the chamber to 1up everyone else when the time is right.
Everyones hyped about the branded phone, but it was the chip that mattered and how fast you rushed a product out after you got it.
Sames true now, except size of training run is also a factor.
Last week I gave it a small-sized auth ticket to work on, then stepped away. I came back later that afternoon and found that it had worked for 3+ hours and written 25,000+ lines of code. I skimmed over the code and it looked like a small fix followed by a massive number of additional checks around it, including static analysis tooling.
I gave it to another GPT 5.6 and said "check this code and see if it addresses the ticket". It looked at it and said that 98% of it was garbage and should be thrown away (its own words). I then gave it to Fable, which said it was massively over-engineered. Fable's theory was that the agent implemented the fix first, but then compacted and lost crucial context, forgot what the original task was about, and kept going. After many compaction cycles it was completely lost.
Some people complain that Opus 5 stops before finishing a task. But to me, that behavior is vastly preferable to what GPT 5.6 Sol does.
Explaining it as a difference of effort would explain both.
Maybe research is sufficiently public and simple to reproduce or the next steps of how to improve things are sufficiently obvious to the smart people working on frontier AI.
What are suspicious of? If the timing is similar maybe just everyone already are of similar capabilities and got there at a similar time?
> Anyone else find it weird how within 2 months of Fable releasing all the major labs suddenly had Fable-level models?
It means Anthropic had no real moat and no real lead. Is that weird to you?
My guess is all the commenters (you are the 4th person I’ve seen say this) saying ‘Anthropic has no moat’ haven’t actually used Fable or even Opus 5 yet. Sol is laughable by comparison, and Grok… lol.
I've used plenty of Opus and Fable. Still do.
> Sol is laughable by comparison
Not really, it depends. Sol is better and useful in some areas. Definitely not all.
Fable is gimped just by those "guardrails" that silently downgrades you to Opus 4.8. Not only do you pay extra for Fable but your caching can be easily messed up. It also doesn't just find all the bugs or is bug-free. Sol has spotted lots of Fable issues and vice versa. Fable also costs 2-100x as much.
> I don’t constantly swear at and call stupid
That's not a judge of anything. There are models that may be stupid and you can swear at it, but if they still get the job done for 1/10th the price... maybe that's all you're paying for.
It's not an explanation of why it happens, I am just pointing Fable is not an exception, it has happened with almost every other model release by all these companies over the last 2-3 years.
it used to be snapdragon came out HTC rushed out a janky phone everyone went omg htc is goat, then in the next few weeks and months others would impliment better versions and people would not notice those as much, finally sony would release a polished phone right as the next snapdragon cycle came.
eventually compute gains leveled off and apple won on taste.
nvidia/tpu is the new snapdragon. Anthropic and google both peaked on the first training run on a new tpu cycle.
you should expect amazing things within a few months of each other from everyone with access to chips and willingness to use them on a training run.
We haven't seen willingness from google to do that. So its currently xai,oai,anthropic, and probably soon meta.
But I think we’re discovering that intelligence is about universality, not magnitude. This is analogous to how building a universal Turing machine wasn’t merely a matter of building a calculator that could multiply higher numbers. The difference is that with calculators we consciously theorized about what universal computation would require, then we built one as a step change. Despite it having low memory and slow speeds, the first one built was as theoretically universal as any computer we have today, in terms of the surface of computations it can perform.
With intelligence, it’s turned out to be less discontinuous, which I believe has convinced people that intelligence is a never ending exponential rather than an S curve approaching a horizontal asymptote. I suspect the LLMs we have today are the same kind of thing we will have in 5-10 years, but in 5-10 years we’ll consider them to be fully universal. At that point we’ll still have improvements in tokens per second and volume of context window, but not in capability per token.
You can also run massive amount of LLMs in parallel.
There might be a limit to a normal LLM but not to theo everall system.
One instance of an LLM is the same as another instance, so while you may get more out of it by stacking more of them, I strongly suspect it falls victim to diminishing returns. 100 instances of the same LLM may converge on the same result as 10.
Not in the highly verifiable domains. There you can take it from say 80-90% maj@x to 99% pass@n. Math, some parts of programming and cybersec are examples of highly verifiable domains. (e.g. if you're searching for a linux LPE, that's expensive to search but easy/cheap to verify - just have a token in /root and have the model retrieve that token)
It's worth a shot at least, as a microservices architect I have a bias that we aren't networking these enough, a single main agent session orchestrating multiple subagents is different from multiple main agent sessions with their own subagents coordinating with each other.
Although I'm also not sure about just how much better models can really get with this technique. Ultimately you're still getting the same model with the same training data, which are the important parts. Asking it to pretend to be something feels like it would just put a color filter in front of the conclusion the model has already predicted, or maybe alter the path to the conclusion slightly or pick a less likely answer that it still could've provided normally.
FWIW I mean if I have an AGENTS.md that encodes my software heuristics (use an interface in situations like X, here's how we name variables, etc.) it generates far cleaner code than if I don't.
Edit- mostly pointing out that stacking 10 base models vs. 10 models with sufficiently different base context isn't necessarily the same attention routing. I suppose I was thinking about tasks that don't have a concrete single answer.
Bigger limit and no limit are very different.
LLMs scale well in almost all dimensions. Context window (working memory) can be a bottleneck but for humans you can’t scale it at all.
intelligence is more like polishing a ball smooth than growing the ball to infinity.
For many tasks, it will be smooth enough.
Their big bet is that models are going to keep getting sharply better, not that they're going to quickly reach a plateau of quality that they can then defend.
Gradually the labs will start engineering verifiable sandboxes for wider domains like videogames
This strategy will hit a plateau in about 18 months and then we're back to diminishing returns and incremental progress along other dimensions (like accelerated inference using ASICs)
They already hire and pay people with research titles for creating and solving problems in their fields.
And a lot of labs say that RL can help everywere and has plenty of way to go.
There's a lot of domains where that simply isn't the case (like bio)
While it's not going to give you an "alphago" effect, it is still enough to work at human levels, augmented with the general knowledge of an LLM, together making it super-human.
If you consider a 5-year outlook, it is also a very temporary job unless you're like a specialist neurosurgeon or something, as one of the examples in that article shows:
> The on-again, off-again nature of the work is not just the result of company culture; it stems from the cadence of AI development itself. People across the industry described the pattern. A model builder, like OpenAI or Anthropic, discovers that its model is weak on chemistry, so it pays a data vendor like Mercor or Scale AI to find chemists to make data. The chemists do tasks until there is a sufficient quantity for a batch to go back to the lab, and the job is paused until the lab sees how the data affects the model. Maybe the lab moves forward, but this time, it’s asking for a slightly different type of data. When the job resumes, the vendor discovers the new instructions make the tasks take longer, which means the cost estimate the vendor gave the lab is now wrong, which means the vendor cuts pay or tries to get workers to move faster. The new batch of data is delivered, and the job is paused once more. Maybe the lab changes its data requirements again, discovers it has enough data, and ends the project or decides to go with another vendor entirely. Maybe now the lab wants only organic chemists and everyone without the relevant background gets taken off the project. Next, it’s biology data that’s in demand, or architectural sketches, or K–12 syllabus design.
Having said that, Grok 4.6 (1.5T params) is without a doubt way smaller than Fable, maybe a Fable sized Grok would be Fable level?
I both agree with this but also find most insistence that Model A is the "best" is social hype divorced from capabilities.
When everyone's improvement (or at least, everyone's rate of increase in parameter count) is so rapid, "within 2 months" shouldn't be seen as "near-concurrent".
Mythos became available internally at the end of February, about half a year ago.
asked grok to give a compute estimate for each: - SpaceX / xAI: ~1.4 GW (owned Colossus clusters) - OpenAI: ~2–3 GW (mostly rented/cloud) - Anthropic: ~1.5–2.5 GW (multi-cloud + xAI lease)
chatgpt estimates a lower: - OpenAI: ~1.5M H100-eq ± ~0.8M - Anthropic: ~1.4M H100-eq ± ~0.7M - SpaceX/xAI: ~0.6M H100-eq ± ~0.3M
but it felt obligated to mention that "for single tightly interconnected NVIDIA training clusters, SpaceX/xAI has been unusually strong."
Makes no sense. At this point, all Western AI companies also engage in distillation. If distillation were such magic, they'd be insane not to.
So why do people have this idea in their heads that it's all some sorta secret sauce they are taking from each other?
So, it's not coincidence when they respond to each others models with something roughly equivalent - because they know what each other are working on.
I think this is the main one. The benchmarks from this are heavily cherry-picked, and they also widely publicised their performance for 4.5 while downplaying the fact the benchmarks were "accidentally" in their training set
DeepSeek V4 Flash 0731 is a distilled version of Fable into the original V4 Flash (announced before Fable), to the point that it also says load bearing and what not.
There are more plausible explanations for why the models are similar - all the labs are buying the same datasets from third parties
Look at deepseek, they improved it just by doing a lot of RL and you can see it from how it behaves. You provide very little information about a task, but since they are trained on similar tasks, they come up with a lot of assumptions and details on their own, because they were trained with such an info during RL.
$60B in SpaceX stock for Cursor was a bargain
Data + compute + being competent and smart enough to ship.
fwiw I don't think these are yet Fable level - the difference tends to get discovered in the long tail of tasks - but they're close enough, they're cheap, and the length of the frontier exclusive window is narrowing
Not if you go by financial fundamentals. All of Space X only has around $18B in sales.
But I think the other reason you didn't mention is the timing of new compute coming online. Compute is the major factor limiting the training of these models and new datacenter investments are bearing fruit at around the same time.
All you need to have Fable-level AI is to announce it, and have enough fans shift from insisting that model Y is the best now, way better than model X.
Couldn't be further from the truth. The models can be tested and statistically evaluated.
I ran a massive Fable max code review on my lone lisp codebase. Now that I have switched to OpenAI, I decided to run an equivalent review using Sol max and compare them. I'm keeping all data so I can thoroughly evaluate their performance in multiple areas such as correctness, rigor, performance, security, maintainability, consistency, among others.
Fable pass is 100% done and I'm around 70% done with the Sol pass. Preliminary results are already becoming clear: Sol is capable of reproducing around 70% to 90% of Fable's performance. Haven't tested open weight models but I'd wager they have the same performance as Sol if not lower.
It seems Fable is still king, I'm afraid. It's undeniable that OpenAI is providing huge value here: up to 90% Fable performance at multiple times the usage on a subscription than what Anthropic offers us is a phenomenal deal. However, if one desires the best model, to me it looks like Fable is still it.
You'll forgive me if I remain unconvinced.
I'm just saying it's not wise to simply put all these models in the same bucket and say any differences are due to vibes or hype. They are clearly different. We can and should scrutinize the testing methodology but it's not exactly fair to just ignore the results.
I don't intend for my benchmark to be private. The core component of my test is my parallel code review skill which is already on my GitHub. I'll be publishing the results on my website when it's done. Anyone could take the skill and reproduce the test using multiple models against any codebase out there, then analyse the depth of each model's findings.
It (mythos) was first made public in April so it's not a surprise that others would catch up, though.
Some people thought this. Some people didn't. Some people thought it was a step backwards. We don't have a solid ground-truth way of estimating this.
Do you have any links to credible claims or independent benchmarks that found they were a step down? Or a specific task that worked worse for you?
My private benchmark tasks, and independent evaluators I've seen all overwhelmingly showed improvement.
Every model released for the past four years has had claims on the internet of getting worse. But transcripts are permanent so it should be easy to give a side by side of an earlier task that is now worse. I don't ever see people do that. Instead I see that every single task on a computer that is verifiable is now night-and-day better.
I'm genuinely curious if you've used them yourself or you're judging this based on internet commentary?
I think one evidence is that the US has more than 5x the compute of China. With that difference in training speed, it should be impossible for Chinese models to close the gap that easily. It's also very unlikely that they sell the same public models to their private customers (military etc). We also know they talk about "unpublished internal models" for things like the last HuggingFace hacking incident. So it's not a bad theory.
How could we really know how much "compute China has" in reality? Is it possible that whatever estimates people has come up with for both China and the US might not be 100% accurate?
I suspect that the models we don’t see are decidedly better than the models we do see.
(On mobile so can't search, but this was yesterday:)
> "Oracle was providing a staggering 22.6 percent of China's known A.I. computing power"
In Malaysia, etc.
They have a big advantage in that they can directly distill from frontier models.
anyone with access to capital can produce frotier model. hell you can just ask chatgpt how to create a fontier model. recipe is not a secret despite what these 'labs' pretend
If this is the case, makes sense that frontier labs with similar access to compute driven by funding on same order of scale can produce improvement largely on similar pace
Frontier model release cycles generally take around 6-8 months anyway. OpenAI and xAI (or however you spell it, branding almost as bad as X/itter) were probably working on their next generation of models already, and Anthropic just beat them 2 months to this release.
You also say "near-concurrent release of the same jump" - but 2 months isn't "near-concurrent", it's a full quarter of the normal release cycle.
I don't think that the other explanations you gave are implausible, though - for both human circulation and distillation, you can apply those during a training and development run (with reduced effectiveness). Reasonable to imagine those as bumping them up another few points to bring competitors from "a little below Fable" to "around Fable".
Anthropic finished a new pre-training run, Opus-sized models got enough of a jump they could have released Fable as Opus 5... but the economics of Opus models weren't where they wanted.
Being the masters of distribution that they are, instead of announcing a massive price hike, they just introduced a new tier and promoted Sonnet-sized models to Opus.
That's why every Opus after 4.6 has had such mixed feedback: smaller model with more RL can only make up so much ground, especially on vibes (which are hard-to-impossible to build a reward for)
(I mention all of this because if they'd just released Opus 5, no one would be asking "why is it a few months later everyone caught up to the latest release"... that's always how it works)
The moral of the story? People work in parallel on the same goals, they build on best practice, or sometimes just need to see something is possible (reusable rockets). Having achievements cluster like this is normal and expected.
So what you experience as a "near simultaneous" release is just their decision of when to peel off a release from their current set of in-training models, likely based on how they perceive market and regulatory conditions. They likely see a competitor release and then baseline what they should release based on that and it takes a month or two for them to package it up and push it out the door.
What I can imagine is that for some of the labs, they are being forced to publish models closer and closer to the frontier of what they have in training. Effectively, "falling behind" is your forward pipeline shrinking. Google ran out of forward pipeline. So far Anthropic and OpenAI didn't - but probably, one is shrinking.
different training makes a different version (agent, info sec etc).
OpenAI has the Doug/Astro pretrain coming up next.
So Model N/N+1 might literally have had the exact the same pretraining run and only differ on how much/what kind of postraining they got
Everyone in silicon valley has a cousin who's supposedly seen Anthropic's new unreleased model that changes everything forever. I distinctly recall sitting in a work meeting where a coworker was insisting GPT4 was AGI that anthropic was too scared to release. It's amazing to me that this playbook is still working at least somewhat on folks.
Where: E = Efficiency, and efficiency gains come from quality of data, quality of algorithms. C = Compute (Size of model, flops of train run)
So a better company can train a bigger and better model with less required compute which let's anthropic get there first. If another company does the same thing with a worse: model architecture, kernel, optimizer, etc... They will get there as well if they just run there train run with more flops for longer
Mythos was actually ready about 6 months ago. So if you have 6 months later or hardware setup and time to train you can get a lot done.
Chinese labs must follow similar trajectories plus their specific efficiency improvements. That also explains the jump from DeepSeek 4 performance in April and July releases. They both use the same pre-trained model as well.
I also stated recently (in informal conversation), based on the performance posted, that said variability was only applied to specific fields of information.
So allow me to make a more provable prediction:
There will be another significant jump related to full field converage, followed by another and from there (we'll call this v3), it will then be capable of automating ASI.
1) AI researchers talk and change companies often, so techniques circulate. This feels implausible because training and shipping a new model ought to take longer than 2 months?
2) Distillation - also implausible for the reason above.
3) Benchmark hacking. AI companies have ways they can dial up performance artificially, and will reach for that to maintain the appearance of parity.
Other reasons?
Edit: Most replies are ignoring timing. It's the near-concurrent release of the same jump in capability that I find suspicious; not the fact that labs can catch up eventually.