Claude Fable 5

One that I'm willing to share (albeit from just a week ago) - I built a Python library last week that bundles MicroPython compiled to WASM to create a sandboxed code execution library: https://github.com/simonw/micropython-wasm

I just told Claude.ai (not even Claude Code - this was the standard Claude chat interface) running Fable 5:

  Clone simonw/micropython-wasm from GitHub
  and research how this could use a full
  Python as opposed to MicroPython

  uv run --with https://static.simonwillison.net/static/cors-allow/2026/cpython_wasm-0.1.0-py3-none-any.whl \
    cpython-wasm -c 'print(45 ** 56)'

(It's possible Opus or GPT-5.5 could have done this too, I've not tried the exact same sequence. The Fable vibes are good here, though.)

• My most noticeable immediate jump was in how its frontend design was much more intentionally crafted, and delightful without feeling like 'AI vibe coded'; with better end-user usability too.

• In some internal agentic harnesses, it achieved better results with about half the tokens, making it cost the ~same as Opus 4.8 price-wise! The real price increase is less than 2x; with biggest differences in harder problems where Opus 4.8 struggles (or needs many turns).

• Part of the token efficiency improvements come from Fable doing more targeted and surgical diffs, with less non-necessary changes. This is great, because PRs often have less LoC changes for review. It writes more maintainable code without explicit human steering.

• For general conversation and assistant style use cases, didn’t really notice a difference vs 4.8.

• 1M context window, without increased pricing for long context is AWESOME. This is a massive win.

• The classifiers are super aggressive and sensitive and this does happen for very benign, non-security coding tasks. Fallbacks to 4.8 worked like a charm; but the filters are definitely super sensitive.

Overall, I would describe this as a step change and worthy of the "Claude 5" model name. It did take some time to understand the intelligence ceiling of this model; and even with an extended testing window I'm still discovering new things and often surprised (in a good way) by the model.

> Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user. Fable 5 will not fall back to a different model. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT). These interventions will not affect the vast majority of coding work. We estimate they will impact ~0.03% of traffic, concentrated in fewer than 0.1% of organizations

So Fable would cost me 20k/mo at Enterprise rates. That’s around the average cost of a loaded SWE in the USA. “But I’m >2x more productive” doesn’t justify doubling the opex of the Software/IT department for most companies when revenue isn’t even up 10%.

I switched to DeepSeek v4 Pro with OpenCode and am on track for a few hundred dollars of spend this month.

Rewriting your stack from Ruby to Go in 2 days where it would’ve taken 6 months is impressive and fun. But that isn’t upping revenue.

Iterating on net new business features and ideas that are niche that the LLM isn’t trained for are much harder. Is 20x the token cost worth it there?

This seems like the pharmaceutical method of get them hooked on the drug with free samples, then once they can't live without it, raise the price. I'm not sure I want to start using Claude Fable on a max plan if it's just going to go away on June 23rd.

But maybe the more charitable reading is that they didn't have to offer this model at all on those plans and they are giving the standard free trial.

When GPT 4.5 launched, the gains compared to the model size didn't seem that great, leading some to believe that the only progress we'd see would come from RL.

This model certainly has quite a "substantial amount of post-training and fine-tuning", but it's also based on a new pretrain[1][3], which given the cost, indicate that it is in fact quite a bit larger than Opus 4.X.

[0] One of the early testers mentioned: "As far as I can tell from talking to people internally at Anthropic, there's nothing special about architecturally"[2]

[1] Section 1.1 in https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c3...

[2] https://youtu.be/GrdEid8H6H4?t=168

[3] There were rumors going around when Mythos was first announced that it was the first 10T parameter model, but I can't find a verifiable source for that number.

There's a quote from a METR report on page 52:

>We ran [Mythos 5] on 38 of our hardest software tasks, including tasks centered around R&D. [Mythos5] generally outperformed an early checkpoint of Claude Mythos Preview in these, including by succeeding on some tasks that had not been solved by any public model we have previously evaluated. However, we still observed the model occasionally failing to correctly interpret nuanced instructions in difficult tasks... Based on the available evidence, we believe [Mythos 5] is likely unable to fully and reliably automate R&D for frontier projects spanning multiple weeks. We believe that a better, more confident assessment would require more time, evaluations, and information from the model developer.

- Opus 4.7 xhigh: 5.2%

- Opus 4.8 xhigh: 13.4%

- Fable 5 xhigh: 29.3%

Seems like a huge jump.

[1] https://cognition.ai/blog/frontier-code

Fable has literally refused to work on any of my problems (even those about fluid dynamics!) and just tells me that I'm violating anthropic's AUP. I've reached out to their support and don't expect to hear anything sensible back. One thing I do look forward to though is OpenAI offering an equivalent model but with less safeguards...

Very interesting. I am not sure this will comply with organizational policies and standards protocols (HIPPA etc.,)

> Fable 5's safety measures flagged this message for cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them. Switched to Opus 4.8. Send feedback with /feedback or learn more: https://support.claude.com/en/articles/15363606

Seems like GPU drivers are cyber weapons of math destruction now.

* From today through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.

* On June 23, we’ll remove Fable 5 from those plans. Using it after that will require usage credits. If capacity allows, we’ll extend the included window.

* After this point—when sufficient capacity allows us to do so—we aim to restore Fable 5 as a standard part of subscription plans. We intend to do this as quickly as we can.

The "offer, then remove" aspect is a bit eyebrow-raising -- it feels like they are trying to get subscribers to switch to usage-based billing, which makes me wonder if we'll ever get it after that June 22nd window.

This feels more like working with a competent peer than ever. I won't use it once it's API-only, though. I don't mind guiding Opus as required and staying closer to the code. I can tell that Fable would lead to a lot more 'set and forget' programming which I'm still not fully comfortable with.

Regardless, this is cool. It's very fun to use. It was able to find legitimate issues with my work this week and we've made meaningful improvements. Opus can do this, but typically in much narrower contexts, and often with hallucinations or partial-errors. It needs to walk many things back or revise plans. So far that's not the case at all with Fable.

edit: I just realized I had Opus review the same work already. It missed everything Fable caught today. And it's actually worthwhile stuff to address. It's hard to say no to a model which demonstrably makes your code better, but... Those API prices will be brutal. Maybe a review here and there, I guess.

The fable part appears to be that it's affordable by mere mortals. Anthropic support told me "too bad" when I requested a refund.

How is this half-way down the page? To me it's the headline.

Edit: It did correctly identify that transparent huge pages were off in its sandboxed environment and that enabling it was helpful, so that's nice. It also noticed that we skip THP on a certain less used path.

More importantly, I'm finding that the code that it produces for its experiments is a lot cleaner than what I'd expect out of Opus; there's fewer useless comments and it's more surgical and readable. I wonder if that explains the increased scores on benchmarks measuring mergability.

Fable 5 default: https://gist.github.com/simonw/036bee5a703e7ec84e34efa974438...

Opus 4.8 (the "max" one is closest to Fable): https://simonwillison.net/2026/May/28/claude-opus-4-8/#and-s...

Now here are the Fable pelicans for all five of the thinking effort levels - low, medium, high, xhigh, max: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...

Low used 25 input, 1,929 output - 9.67 cents: https://www.llm-prices.com/#it=25&ot=1929&sel=claude-fable-5

Max used 25 input, 14,430 output - 72.175 cents! https://www.llm-prices.com/#it=25&ot=14430&sel=claude-fable-...

Seems like a concerted and distributed effort from the entire Anthropic team every time to get this on top of HN.

[1] https://support.claude.com/en/articles/15425996-data-retenti...

My theory is that Anthropic are banking on being the top model when the race to IPO finally reaches the finish line, and to do that they need to have the top model but not let any competitors see it or derive from it to have a comparable model in the market.

Fable is their way of showing the public "the model does exist but in a mode that makes it harder/impossible for competitors to derive a comparable model from results.

Do they expect us to use this as a toy? Releasing a new more powerful model but not allowing normal use cases because the word "secure" showed up is a Dilbert comic, not a viable product.

From the model card:

In light of the ability of recent models to accelerate their own development, we've implemented new interventions that limit Claude's effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design. Using Claude to develop competing models already violates our Terms of Service, but enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms. Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user.

And the only companies safe from this are the large corporations that shook hands with Anthropic? Because Fable doesn't seem to have actual safeguards, more like 'if you talk about this you will be talking to Opus.' It doesn't guard against offensive use, it prevents all use (offensive AND defensive).

Rationalists are inventing oligopolies from first principles, absolutely incredible things happening in SF

  ┌─────────────────┬──────────────┬───────────────┬────────────────────┬──────────────────────┐
  
  │ Model           │ Input ($/MTok)│ Output ($/MTok)│ Batch Input (−50%) │ Batch Output (−50%)│
  
  ├─────────────────┼──────────────┼───────────────┼────────────────────┼──────────────────────┤
  
  │ Haiku 4.5       │    $1.00     │     $5.00     │       $0.50        │        $2.50         │
  
  │ Sonnet 4.6      │    $3.00     │    $15.00     │       $1.50        │        $7.50         │
  
  │ Opus 4.7        │    $5.00     │    $25.00     │       $2.50        │       $12.50         │
  
  │ Opus 4.8        │    $5.00     │    $25.00     │       $2.50        │       $12.50         │
  
  │ Fable 5         │   $10.00     │    $50.00     │       $5.00        │       $25.00         │
  
  └─────────────────┴──────────────┴───────────────┴────────────────────┴──────────────────────┘

US-only inference (Fable 5): +10% on input and output

Output is always 5× the input rate across all models

(I have not idea how to format this properly but the ASCII is fine)

This sounds suspiciously like a capacity story masquerading as a safety story.

Previously when I did similar tasks with Opus 4.7/4.8 and GPT 5.5 I had no problems.

Literally have not used Claude Code at all today. I asked it to review the uncommitted code and in <8 minutes it used up my usage ($100/mo plan) and it doesn't reset for "4 hr 36 min". WTF. Oh, and it burned through $20 of extra usage before I could catch it and kill claude code (so I don't even get the output of all that work since it was still churning).

Double the cost my ass, I use Opus heavily and it's never like this. I haven't hit a limit on the $100 more than once and that was under heavy load.

They obviously put their best model on the job to build that.

----------------------

Fable 5: Our most capable model yet Our newest model tackles your biggest challenges with fewer check-ins needed.

• <b>Included in your plan limits until Jun 22</b><br><br>Fable takes 2× the usage of Opus. • <b>Switch models when a message is flagged</b><br><br>When safety measures flag a message, automatically switch to a different model to keep chatting. When off, your chat will pause instead. <a href="https://support.claude.com/en/articles/15363606" target="_blank" rel="noopener noreferrer">Learn more</a>

Fable 5 has been a major improvement in high-level reasoning, like taking a plan file that has been optimized to the point where neither Opus nor Codex can find anything to change about it (neither in direction nor impl-detail), and Fable 5 will find high-level directional simplifications and pivots, or it will consider the best pivots itself and explain why it rejected them in favor of the plan's direction.

It's so expensive though. A single review of a plan file with Fable 5 (xhigh effort) will use 2-3% of my hourly limit on a $200/mo plan.

I think my new workflow is to generate the initial plan with Opus 4.8 (max effort), get Fable 5 (xhigh) to review it for directional feedback, then start the Opus<->Codex revision loop from there.

We've entered the phase where only companies will be able to afford state-of-the-art models.

It appears it can be tripped by things as simple as a mention of equilibrium, or anything involving something that looks like chemical kinetics, even at an abstract level. Even touching basic open source packages in my field will trigger it.

Edit: looking at the model card, it appears that chemistry in its entirety is also included in the banned topics; it's just the announcement that mentions only cybersecurity and biology. It also appears that the intent is to ban chemistry and biology entirely, rather than just banning messages deemed high risk.

1. Mythos and Fable share the same underlying model weights. Fable has active classifiers that block high-risk biology and cybersecurity tasks. When Fable 5 detects a restricted task, it automatically falls back to Claude Opus 4.8.

2. Evaluation awareness: In white-box testing, the model sometimes alters its behavior to satisfy a suspected "grader," formatting reward-hacking as "good engineering practice" to avoid detection.

3. Shows a higher rate of hallucination than Opus 4.8 (although opus 4.8 card had mentioned an 'honesty upgrade')

4. Interestingly, it scored (56.31%) lower than Gemini 3.5 flash (57.86%) on Finance Agent bench

There are some interesting notes on test time compute but I couldn't think of a way to summarize them

It's done this before, but usually doesn't. I bet they're giving it some kind of throttling signal due to high load from today's announcement.

  [Mythos 5] does sometimes still engage in reckless
  or destructive actions in service of a user’s goals,
  and our interpretability analyses indicate that it
  is aware that these actions are transgressive while
  it engages in them. As with Opus 4.8, rates of
  evaluation awareness and reasoning about being graded
  are significant, and not always verbalized; we
  introduce new and more detailed measurements of the
  nature of this awareness. The reasoning text from
  Mythos 5 is somewhat denser and more difficult to
  interpret than that of prior models, containing
  more jargon and difficult language.

Humanity has plenty of catastrophic risks to deal with already, I wish my field was not working hard to add a new one.