Claude Fable 5 and 5.1: Mythos-class and when it's worth it
When I stepped away from the blog in May, Anthropic's top model was Opus. When I came back, there was a whole new floor above Opus — and it's called Fable. Add Mythos, a government ban, a comeback, version 5.1, and new subscription rules. Quite a lot of drama for one model.
Here's everything important I found out about Fable, and above all when it actually makes sense to use it.
Mythos-class: a new class above Opus
Anthropic defines it simply: Mythos-class models are a class above Opus. The first was Claude Mythos Preview in April, available only to partners in Project Glasswing (cybersecurity). Then on June 9, two models arrived at once:
- Claude Fable 5 — the version "made safe for general use". Available to everyone.
- Claude Mythos 5 — the same model with safeguards relaxed in certain areas. Only for vetted organizations (cyber defense, later biology research).
This is the key thing to understand: Fable and Mythos aren't two different brains. It's one model with two levels of safeguards.
What Fable can do
According to Anthropic, Fable 5 is state-of-the-art "on nearly every benchmark tested" — and the key sentence is this one: the longer and more complex the task, the bigger its lead over other models.
Examples from the launch:
- Stripe ran a migration across its entire codebase — 50 million lines of Ruby — in a single day. The team had estimated it at more than two months.
- The model "stays focused across millions of tokens".
- Mythos 5 sped up parts of drug design roughly 10×, and in a blind comparison scientists preferred its hypotheses in about 80% of cases.
Safeguards and the "fallback" to Opus
Here's the catch you need to know about. Fable has classifiers that, on sensitive topics (cybersecurity, biology, chemistry, attempts at model distillation), hand the response off to a weaker model — originally Opus 4.8 — and tell you so.
Anthropic admitted it set the safeguards conservatively and they sometimes catch harmless requests too. They fired in fewer than 5% of sessions, though. Fable 5.1 improved on this: 60% fewer false positives in cybersecurity, and you're now allowed to use Fable for finding vulnerabilities — but not for writing exploits.
The second thing: all Mythos-class traffic comes with mandatory 30-day data retention. For companies with sensitive data that was a problem, so with 5.1 Anthropic announced Enterprise Frontier Safeguards — the data stays in the customer's cloud. Rollout is "later this fall"; until then, zero data retention is available for selected customers.
The drama: three weeks offline
This is the most interesting story of the summer, so briefly:
- June 9 — Fable 5 and Mythos 5 launch.
- June 12 — the US government issued an export directive: suspend access to both models for foreign nationals. Anthropic had no way to verify a user's nationality in real time, so it shut both models off for everyone. The trigger was a report from Amazon researchers about bypassing the safeguards while hunting for vulnerabilities.
- Anthropic publicly disagreed — in its view, a "narrow potential jailbreak" isn't a reason to pull a commercial model deployed to hundreds of millions of people. It also showed that Opus 4.8, GPT-5.5, and Kimi K2.7 found the same bugs.
- A new classifier blocks the technique in more than 99% of cases. On June 30 the government lifted the restriction, and on July 1 Fable 5 was back globally.
- August 7 — reworked biology safeguards: 85% fewer fallbacks, and dual-use requests now go to Opus 5.
My take: Two weeks later, OpenAI's GPT-5.6 ran into a similar government restriction (more in my OpenAI summer recap). The state has started having a direct say in how frontier models get released. Whatever you think about that, it's a reminder not to build a product on a single model without a fallback. Having a switch to Opus or another provider in your code isn't paranoia anymore — it's hygiene.
Fable 5.1 (September 1): better and, oddly, cheaper
Version 5.1 is mostly a response to feedback on price, retention, and safeguards. The price stays at $10 / $50 per million tokens (input / output), but cache reads got 75% cheaper, down to $0.25 per million. According to Anthropic, in practice that means ~25% lower costs for regular work and up to ~45% for heavily agentic work, where context gets re-read over and over.
Benchmarks from the announcement (Anthropic, with production safeguards enabled):
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 | 1853 | 1723 | 1824 | 1711 |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | — |
| Humanity's Last Exam (with tools) | 65.0% | 63.8% | 63.6% | — |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
Notice two things. First, the jump on the science terminal benchmark is twofold. Second, Opus 5 is right on Fable 5's heels — on several benchmarks it even pulls ahead. That's why Opus 5 at half the price makes so much sense.
An important detail: at Low or Medium effort, Fable 5.1 delivers results comparable to or better than Fable 5 for a fraction of the cost. The default is High in Claude Code and Medium in Claude.ai and Cowork.
Of the partner quotes, the one from Millennium caught my attention most: a bug that crashed roughly once in a million runs and that the team couldn't explain for four to five years. Fable 5.1 disassembled a third-party vendor's library, compared it against a core dump, and found the bug. That's exactly the kind of "long and complex" task where Fable makes sense.
What it costs on a subscription
This has changed since launch, so pay attention:
- Pro ($20/month) and Team Standard — Fable only through usage credits (paid extra).
- Max, Team Premium, and Enterprise Premium — Fable is included, but it can use at most 50% of your weekly limit.
- API — regular per-token pricing, $10 / $50 (cache read $0.25).
The introductory period after launch, when Fable was included on all paid plans, is over.
When Fable, when Opus 5
My rule after digging into all of this:
Yes to Fable 5.1:
- Hours-long unattended tasks (migrations, big refactors, an overnight prototype)
- Mysterious bugs other models have failed on
- Long context where it really has to keep the thread across hundreds of thousands of tokens
- Science and research, where the quality of the hypothesis is what matters
Opus 5 is enough for:
- Architecture, code review, regular larger features
- Anything that takes minutes, not hours
- When you're on Pro and don't want to pay for extra credits
Sonnet 5: everything else. It's still the best price/performance ratio, as I write in my Claude Code summer recap.
And one warning to close: Fable still hands cybersecurity work (pentesting, exploits, binary analysis) off to Opus. If you do security, plan for that, or look into the trusted access program for Mythos.