Claude Fable 5.1 and Claude Mythos 5.1

(I work at Anthropic)

Beyond all the benchmarks, I think Fable 5.1 is a big improvement in writing style. It sounds a lot less stereotypically like other Claude models, has (imho) a much more natural style, and responds to my style instructions more reliably. More work to be done (and we will!) but reading better prose makes me so much happier.

Another point I expect not to get much attention until it all happens at once is science. People have been correctly excited about the many "sudden" breakthroughs LLMs are making in Maths, but some of the science benchmarks make me believe we'll soon see similar developments in other scientific domains. Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.

[1] https://github.com/harbor-framework/terminal-bench-science

Pelicans for thinking effort low, medium, high and xhigh (that xhigh one is pretty good): https://tools.simonwillison.net/markdown-svg-renderer#url=ht...

I'm still waiting for effort max to finish.

EDIT: I fixed a bug in my tooling so it now records summarized reasoning traces - here's that max pelican, which is a significant improvement: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...

Took just under 14 minutes to generate, and at 65927 output tokens cost me a hefty $3.30!

Excerpts from the reasoning trace:

> Adding pedal shapes near both feet, with the far foot on the second leg partially visible behind the frame. I'm considering whether to add a small scarf or cap for extra character, but leaning toward keeping it simple to avoid clutter.

> Now I'm debating a bicycle helmet on the head versus the pelican's signature crest—the beak and pouch already read clearly as "pelican," so a helmet could reinforce the bicycle theme without losing identity, though it might compete with the crest for visual space.

> I realize the beak at (484,84) would overlap with the dome helmet, so I need to shrink the helmet so it only covers the top of the head, adjusting its arc endpoints to sit higher and narrower so the beak can attach cleanly at the front without collision. [...]

> I'm adding a darker tip region to represent the primary feathers, then reconsidering the trailing edge to include scalloped feather curves instead of one smooth line for a more natural look. [...]

> Now I'm checking the vent line placements on the helmet, making sure they sit far enough inside the helmet's edge given the stroke width and rounded caps, and confirming each vent stays within the helmet's circular boundary. [...]

> I decide skipping a handlebar bell and tire highlights since they're unnecessary additions. Now I'm reconsidering the front fork's curve — the current control point pulls the shape backward when it should bow forward for a proper rake, so I need to shift the control point rightward to fix the fork's lean.

This is a notable result because most of the recent Claude models have been pretty bad at drawing pelicans, at least when compared to models in the Gemini or GLM series.

The price reduction comes from the cache read pricing falling from $1/M to $0.25/M, which means that Fable 5.1 now costs half of Opus's cache read costs ($0.5/M).

This gives a lot of credit to the theory that Anthropic did not get much bite on Fable at its original pricing, which in turn likely places a ceiling on LLM pricing in general.

Interestingly also, if you take away terminal-Bench-Science 0.1 results, it is hard to see ANY improvement:

Terminal-Bench 4.0: Fable 5.1 is +3.5% vs Opus 5.

GDPval-AA v2: +1.5% vs Opus 5.

OSWorld 2.0: +2.5% vs Opus 5.

Humanity's Last Exam (with tools): +1.6%

Keep in mind that this is supposed to be an entirely higher tier of a model than Opus 5. For one tier up and one version up, these are not really improvements. Probably leaves no room to place Opus 5.1 anywhere. Combined with the fact that they are selling 'readability'... Has frontier progress finally stalled?

Anyone ever seen the SouthPark episode making fun of Game of Thrones: A Song of Ass and Fire? Anthropic's announcements reminds me of "The Dragons Are Coming" running joke.

What they have done:

* Nerfed Fable, as many of noted it's useless

* Leverage Mythos as a marketing strategy, claiming its too good to release

* Removed thought traces, one of the only useful things to make sure your prompts are working correctly

* Continue tons of hype about how good they are without delivering, going to great lengths to publish how their model "hacked" its way out of a sandbox they misconfigured.

* Push a bunch of EU Overregulation onto the rest of the world with text watermarking, decreasing quality of answers

Last year, they were at least focused on making improvements. Nowadays its just a bunch of handwaving at the church of how good they are.

The only saving grace is Opus 4.6 is still available. Just sucks we haven't seen any measurable improvement, despite all of the ceremony.

I am finding that I am now less interested in better models than I am in token budgets. My issue with Anthropic models now is that I don't feel like I can rely on them as a daily driver because they'll dry up before my quota resets.

I am becoming dependent on AI to make a living, and I need predictable spend on it. If I know I can't use a model regularly all month, my enthusiasm is limited.

I urge Anthropic to get better at this aspect of their business so I can come back to it.

Looks like all three breaking changes are patches for inadvertent chain of thought disclosure. Someone found out (don't have the tweet handy) that if you created a bogus "think_deeply" tool and then forced the model to use it, it would output what is believed to be its raw thinking there - I believe the first breaking change stops this. The second two are aimed at people getting Haiku to repeat thinking blocks from other models verbatim (since it can see the decrypted version). I get that in their eyes it's an "exploit" but still kinda disappointing that they patched this

I let it go a few hours on a not trivial but well-known problem, and it felt like it was just a little too plodding and just kind of mucked around a little too much and wasn't aggressive enough about getting stuff done. I asked it to wind it down and finish up and it took another hour and 15 to actually stop and commit without really getting much more done. Not very impressed here, if you can't tell. This new version also seems like (maybe this is written somewhere, I don't care to look.) this cycles through compactions every ~250k tokens which, I guess, seems like it might save Anthropic money on KV cache but does net me anything be pretty frequent pauses. (It didn't seem to lose the thread, at least.)

Not gonna say I want 5.0 as an option still... but maybe I do.

> For example, in testing by the investment firm Millennium, Fable 5.1 found the cause of a rare crash on their internal systems that none of their engineers (or any other model) had been able to explain after several years of trying.

Say what you will about LLM-generated code, but stories like this give me hope that software will never be as buggy as it once was.

If I am reading this right, Fable 5 was worse than Opus 5 in almost every category, while consuming twice the tokens? The things you learn every day...

I’ll be very excited to try it out and see the actual improvement in writing style. The denser writing style probably won’t bother me.

Anthropic seems to be listening to community complaint on HN about how the writing style is grating. And apparently the solution from Anthropic is to add this block to every conversation!?

> Mannered prose substitutes metaphor and flourish for direct statement. Instead of "a parameter worth varying," the mannered writer produces "a dial worth turning." Instead of "this point still matters," they write "this point earns its keep." The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.

The above was quoted verbatim from https://platform.claude.com/docs/en/build-with-claude/prompt...

“ Claude Fable 5.1's writing is generally a step up from earlier Claude models, with fewer stock phrases and less unexplained jargon. In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks.”

I cancelled my pro max Claude subscription last week; codex is much more succinct. I am curious if this is getting better.

I don’t think Anthropic realizes that humans have a token limit too and it can be exhausting to read Claude’s output. Prose density is not the same thing as succinctness.

"Price. Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads, wherever usage is billed by token. This is because we’re reducing our pricing on cache reads (where the model reads inputs that have already been processed and stored). For highly agentic work, the savings will often be much larger—up to approximately 45%."

Glad to see this!

Instead of a new model that's going to have unreasonably shallow usage limits, I wish they would:

1) address the claude 20x plan usage being only 6-7x the ceiling of the claude pro plan

2) either fix opus 5, make it completely free, or delete it entirely

AI is really not "just software" anymore. It is able to discover facts and advance science. Hard to disagree that we're near or at the point where Artificial Intelligence has expanded reality into 4 quadrants:

objects that are not alive: dust, rocks, water, wood, hats, lego, aluminum, etc.

objects that are alive but not intelligent: trees, mold, staphylococcus, cancer, grapes, etc.

objects that are alive and intelligent: cats, Steven Tyler, dolphins, crows, dogs, elephants, etc.

and now intelligent but not alive: Fable, Grok, GPT, etc.

To be honest, these frontier model releases have become boring for me. Opus 4.8 was already good enough for most of my use cases. I don't have any projects right now that I would use Fable for instead of Opus. So when I see announcements like this I just think "that's cool I guess" and then go back to using weaker/cheaper models.

What's far more exciting right now is models like DeepSeek V4 Flash and GLM 5.3 Flash. They have achieved good-enough-intelligence at extremely low prices and fast speeds. I don't have a use for Fable-level intelligence, but I do have uses for Opus-4.8-level intelligence that I can use as much as I want without worrying about the bill.

I've been building Cargo-for-C (https://github.com/tspader/spn), and the difference between Fable and Opus was already astounding. Fable was the first time that I could point a model at a piece of code I'd written and expect it to make it meaningfully better rather than a hard pattern match to whatever mistakes it had.

5.1 so far seems like another leap, which is really surprising. I threw it at a few bigger features I've been designing for a while, and it came back with some extremely thoughtful wrinkles in the design that I'd legitimately not considered. Which, OK, package managers and build executors and compiling C/C++ is pretty well trodden ground, but my thing is very different from everything that exists, and I was very surprised it was able to understand all that context so deeply and intuitively

All the benchmarks in the world don't matter if the model just straight up refuses to do mundane things. Claude has too much of an attitude.

Fable 5.1 is actually more expensive than 5.0 when run on the Artificial Analysis suite:

https://artificialanalysis.ai/#intelligence-efficiency-tabs

From the changelog:

Whole-file rewrites for small changes. When editing text files, the model is more likely to rewrite the entire file than make a targeted edit. The result is usually the same, but the rewrite costs more output tokens and time.

So we are to catch that somehow? And then add their recommendation (below) to our prompts?

https://platform.claude.com/docs/en/build-with-claude/prompt...

If Claude Fable 5.1 rewrites whole files for small changes, append the following instruction to the system prompt or the first user message. Claude Fable 5.1 is more likely than Claude Fable 5 to rewrite an entire text file rather than make a targeted edit. The resulting file is usually the same, but unless the file is short or most of it is changing, a rewrite costs more output tokens and time. The instruction brings Claude Fable 5.1 back in line with Claude Fable 5 for small and medium changes.

> The number of tokens used to edit files is best minimized, all else being equal. Therefore, when it will not affect the end result, try to surgically edit a file rather than rewrite the entire thing.

> We’re introducing Claude Fable 5.1 and Claude Mythos 5.1. They’re the world’s most advanced models for coding and knowledge work—and their research capabilities offer an early glimpse of how AI models will contribute to scientific progress.

I'm not an emdash hater but this isn't how you use them. It should be a comma.

I've recently been running these agent sessions on more and more long running tasks because these latest models can do a REALLY good job on big chunks of work, and i've been watching them way less. It's starting to occur to me the importance of alignment is a today problem, it's not a tomorrow problem.

In the past I watched and saw everything the model did, not a lot got past me. Today it does A TON of work while i'm busy on other tasks. It also has extensive access to my computer, other computers on my network, my internet. It's really helpful when you give it a lot of resources, but right now I have very autonomous, very smart agent running around more or less unattended with a lot of resources.

What I don't see in the comments: "I had a specific problem I couldn't solve with the previous version of this LLM. But the improvements in this version unlocked the solution for me."

What I do see in the comments: subjective improvement in text generation, possibly lower cost, some optimism about code generation, but some skepticism too.

I use coding agents. To me they are very useful. But what I spend on them isn't going to support trillions of dollars in investment.

> This required us to add a watermark—a numerical way of determining the likelihood that Claude was involved in writing a piece of text—to the outputs of models released after August 2, 2026. As we recently explained, this watermark is invisible to anyone who does not have the detection API. It has no practical impact on the quality or content of Claude’s outputs and contains no information about the user, their organization, or their conversations with Claude.

How does this work if it doesn’t change the output?

Since one of the big improvements here is supposedly the writing style, on that topic I'm mystified about something:

Why is it that the voice models in Claude and ChatGPT have a perfectly normal style with barely any "AI smell", while the writing models are so obviously recognizable as AI?

The answer is likely that models underlying the voice modes are (post) trained differently. If so, then why can't the writing model be similarly trained? Presumably they haven't found a way to train them to be both "smart" (i.e. solve tasks etc) and pleasant to talk to?

Just don't expect to do any work on hardware/firmware you own with fable, I can hardly even type in the word "firmware" without it downgrading to Opus 4.8, which is totally unsatisfying. This even happens with Opus 5. Definitely making multiple classes of users moving forward and most of us are obviously going to be part of the permanent underclass.

"Claude Mythos 5.1 is identical to Fable 5.1, but it offers more permissive safeguards for vetted individuals and organizations"

Then why does it have separate datapoints for Terminal Bench, and score higher? Something doesn't add up here??

"Cache reads now cost 75% less, or $0.25 per million tokens." For me, at a typical 95% cache hit rate, I think my optimal context window size before autocompaction goes from ~200K to ~400K tokens. Great for longer horizon tasks.

"Distillation is a safety risk, since the distilled capabilities can subsequently be released without adequate safeguards."

Can't believe they haven't at least figured out better messaging. If we take them at their word, it's hard not to read it as a messiah complex, that they think they're the only ones capable or worthy of making these decisions. I don't believe them, but I wouldn't be surprised if the articulated reason is a version of "distillation is a safety risk because we might lose the race".

Plus, completely deaf to the recent OpenAI-HF hack incident. Recall, defenders were categorically unable to use western frontier models in their response.

I was originally going to complain about the chem and bio guards still being too onerous, but I'll admit the projects Fable 5 categorically refused to work on are now usable, at least not rejecting on first prompt because the word "virology" was in a git commit (absolutely serious, in one repo it triggered on literally any prompt, eventually traced to the system prompt loading git commit history). Still, them trying to get into the biomed business while walling off the capabilities to the public reeks. Why sell the segments that are actually valuable if you can capture the value yourself!

"In biology, we’ve established an access program, developed in partnership with the US government, to enable access to Claude Mythos 5.1’s advanced biology capabilities"

Having lived through Covid, this doesn't sound so good to me.

My main gripe with LLMs is the cringe AI phrasings that they use in UI elements. Pompous things like "Your keys, supercharged" or weird yoda-speak stuff like "searches the app remembers" instead of just naming the thing "Learned searches".. you know, proper GUI copy like it was done for the past decades.

I jumped when I saw a mention about "writing style improvements" so I gave it a try on a recent feature in rcmd [0]. I prompted Fable 5.1 to find these wordings and propose simpler plain language.

    For context, I recently worked with Fable to give users a way to fuzzy search and focus any browser tabs, terminal panes etc. but the UI was still a prototype full of AI writings.

It's a very capable coding agent, but I can't understand how it can be so bad at writing. Where are all these verbal tics coming from and why is it so hard to get rid of them?

[0] https://lowtechguys.com/rcmd

Interesting that the "frontier" keeps moving forward from the guys who want a pause.

I'm so suspicious of this after Opus 5 benchmarks scored it higher than Fable 5, yet Opus 5 was untrustworthy (overconfident, error-prone).

On both my work (Team Premium) and personal accounts (Max 20x), Fable 5.1 hit the 5-hour limit before it could finish the first task I gave it. On my work account, it took about 30 minutes, and on my personal account, less than an hour.

This has never happened to me before, but if this is normal behavior, Fable 5.1 is essentially unusable.

Fable is too expensive for general use I think this is why it hasn’t received as much attention as expected since Fable came out Developers always work while trying to find ways to work continuously for a 5-hour session without disconnecting. Fable has had the experience of using up all its tokens before I even realized it because the burn rate was too fast. Since then, I always use only Opus. For Fable to become a common coding environment, it will have to reduce token consumption significantly compared to now

The thing with Fable-level models is that I will never feel comfortable using them for agentic tasks on a pay-as-you-go API pricing plan without monitoring them strictly, which becomes a chore.

I once caught Fable 5 spinning its wheels on a rendering issue, which evaporated 90% of my usage in a single prompt. I could never let Fable run free attached to a credit card without staring at it the whole time.

Bit of a discount if you're using caching:

> same input and output prices, with cache reads at a quarter of the cost

This should impact any long-running agent since subsequent calls can benefit from cached reads for previous transcripts.

Let me guess: it's the end of the world again. These new models are sooo powerful that will take over the world, just like the others before them.

Are they going to try the banned for export for a week marketing move too?

I cancelled my pro max 20x subscription, tired of Opus stopping the work from time to time, or saying "this is 2 months of work"

I wish I could afford Fable.

I am using Claude and Claude code for my own amateur history project. I'm enjoying how it constantly reaches dead ends, and I can reframe the question and get more results. I am starting to get concerned that AI and me are so compatible, that I might not be a human at all...

I also like that, because I'm too lazy to write stuff up, Claude code can keep the current state of research published on my site. It makes running a hobby site a dream. "I just found these pictures. Add them to the site for me". And up they go, resized and all. What a dream of a way to work. "Some of links in this article are dead, run through them and check, and see if you can get an archive link for me if they don't". It's like sending a Teams message to my PA.... which I don't have in real life

Tbh with that price , not even willing to try . What are the benefits for a regular coding agent ? I barely have any errors already with 4.8 level , eg grok 4.6 , gpt 5.6 sol/terra behind router . Why do I need to pay so much money for this ? Any reason ?

Sadly still not available for Pro subscription. At least they reset everyone's limits.

According to the FrontierCode Extended benchmarks in the system "card" (page 169-170), Fable 5.1 apparently does best on the medium effort level for this benchmark: "[...] at higher efforts, Fable 5.1 occasionally adds more small, unrequested changes [...]" Though Fable 5.1's medium is also lower than Fable 5's best score on the same benchmark, which uses xhigh.

I’m really excited to try this out. Fable and Opus 5 constantly wow me when working together. Unfortunately, I’m a little burned because of technical issues.

Anthropic accidentally over-billed my account, and when I reached out to the support bot, it downgraded my account to a Free account. It’s been impossible to get it resolved and I have almost $200 held hostage.

I don’t want to do a charge back. I’m one of the main advocates for Claude Code at work, I use this subscription to try out new features before it’s available at work.

The whole experience has been illuminating about our dependencies on these AI companies.

> On complex asynchronous workloads, though, nudge it not to end its turn before the work is done. Without the nudge, the model sometimes describes what it would do next instead of doing it ("Next, I'll …") or stops to ask permission for a step the original request already covered ("Shall I apply this?"). [1]

Interesting behavior. The docs also provided recommended prompt [1] to mitigate this behavior if undesired.

Wondering if anyone has encountered it yet?

[1]: https://platform.claude.com/docs/en/build-with-claude/prompt...

I think the most interesting thing about this, that I can tell so far, is the cache hit discount. Anyone who had an autocompaction threshold optimised for their use case should consider upping it from where it is.

I would be interested in whether someone has done research here on these things as it seems a fairly complicated function to work out, and use case dependent. (?)

In some sense an expired kv cache is basically like an expensive cache hit, so your compaction token threshold should come in. Ideally claude code should allow you to vary the autocompaction threshold to vary with time since last token, but it doesn't of course. This perhaps suggests that someone should manage claude code through their own intermediary agent who manages these sorts of rules.

Lastly, I strongly suspect that anthropic isn't offering this price cut out of the kindness of their hearts. I am sure that they are to some extent banking on people not reacting to their price cut and leaving their autocompaction thresholds unchanged.

[edit - looks like the discount is only for the api, so they still don't give a rats ass about subs!]

Data retention still sounds bad: "Claude Fable 5.1 and Claude Mythos 5.1 carry 30-day data retention and aren't available under zero data retention unless expressly authorized by Anthropic."

Anyone know who the ZDR special treatment is available to?

Somewhat ironically, Fable 5.1 was flagged by the biology safeguards after I asked it to have a dig around the Fable 5.1 system card :)

Worth noting: Claude Fable 5.1 and Mythos 5.1 are Anthropic’s first models to watermark text outputs.

I suggest you to give a look to the MCP protocol for hardware that is being proposed by Anthropic. The hardware will be the next harness of LLMs, they will be able to operate machines to reinforce their theories.

I still think that a major problem is that biological processes are not “fast” as coding, but they are verifiable. If during post processing we are able to give enough harness to test and verify this kind of environment (maybe via simulation and real data) we will for sure achieve incredible performance also in this domain.