Claude Fable 5.1 and Mythos 5.1: What Changed, What It Costs, and Who Should Care

One Model, Two Doors

Three months ago Anthropic shipped Claude Fable 5, had it taken offline three days later under a US export-control directive, and got it back on 1 July. So there was some interest in how they’d handle the sequel.

The answer, on 1 September, was to ship it twice. Claude Fable 5.1 and Claude Mythos 5.1 are the same model. Same weights, same brain. What differs is the safeguard level. Fable 5.1 is generally available on every platform Anthropic serves. Mythos 5.1 only reaches organisations inside trusted-access programmes for cybersecurity and life-sciences work — the places where a model that will walk you through exploit code or a lab protocol is the point rather than the problem.

I can’t think of a precedent. Labs have gated whole models before — OpenAI gates GPT-6 Astra’s cyber capabilities behind its Daybreak programme, and Google runs Gemini 3.8 Flash Cyber through its own limited-access programme. But one model, two safety settings, two names on the price list is new. It’s a quiet admission that “safe enough for everyone” and “useful enough for a penetration tester” are different lines, and that pretending otherwise was costing capability for the people who legitimately needed it.

For most readers the practical upshot is simple. You’re getting Fable 5.1, and Mythos 5.1 is not something you apply for on a whim. It was also reportedly live for some users before the announcement went out, which tells you something about how tidy the rollout was.

What Actually Changed

At low and medium effort, Anthropic’s own framing is “similar or better” than Fable 5. Translated: don’t expect fireworks on a quick email. Push the effort setting up and the gap widens sharply. The gains are concentrated where the model is working for a long time without you — long-running agentic coding, multistep research, and the unglamorous document, spreadsheet and slide work that fills most people’s afternoons.

The trait people notice first is debugging. Fable 5 was already good at fixing the thing you pointed at. Fable 5.1 has a habit of going off to find why the thing broke, and fixing that instead. Root cause rather than band-aid. If you’ve ever watched a model “fix” a failing test by deleting the test, you’ll understand why this matters more than any benchmark.

Claude Code users get a related bonus: around 60% fewer cybersecurity false positives, which in practice means fewer refusals when you’re doing legitimate security work on your own codebase. Context stays at a million tokens. And there’s a small, pleasant quality-of-life change — the model now sends readable progress updates between tool calls, so a long agentic run tells you what it’s doing rather than going silent for four minutes.

What “25% Cheaper” Actually Means

Here’s where the marketing needs translating. The list price hasn’t moved: US$10 per million input tokens, US$50 per million output, exactly as Fable 5. What has changed is the price of cache reads, which has been cut to a quarter.

In plain English: when you give Claude a long document, a system prompt or a codebase and then keep talking to it, most of what it reads each turn is the same material it read last turn. Prompt caching lets it reuse that instead of processing it fresh, and you pay a reduced rate for the reused part. That reduced rate just dropped by 75%.

So the saving depends entirely on how repetitive your usage is. Anthropic’s figures are roughly 25% cheaper for typical workloads and up to 45% cheaper for highly agentic work — the kind where an agent loops through the same large context dozens of times. Fire single questions at it with no shared context and you’ll save nothing. Run agents all day and this is a real line in the budget.

On Claude.ai the picture is by plan. Fable 5.1 is included on Max, Team Premium and the premium Enterprise tier; Pro and Team Standard get it through usage credits. Worth checking which one you’re on before you assume it’s “in” your subscription.

If You Built on Fable 5, Read This

“Same price, drop-in upgrade” is not quite the truth for anyone with a working integration. Four things changed in the API, and two of them will break code that ran happily on Fable 5.

None of this is exotic. It’s an afternoon’s work for a competent developer. But it is work, so test before you change the model string in production.

The June Lesson: Keep a Fallback

I wrote about the Fable 5 suspension at the time and I’m not going to relitigate it. The short version bears repeating, though. On 9 June Anthropic launched its best model. On 12 June a US export-control directive took it offline. On 1 July it came back. Nineteen days. Anyone who rebuilt a process around it that first weekend spent most of June explaining to their boss why the new thing had stopped.

Fable 5.1 arrives with a second door marked Mythos and a fresh round of regulatory attention on the cyber question across the whole industry. I’m not predicting a repeat. I’m saying the sensible architecture hasn’t changed: keep your prompts portable (the RCTF discipline exists for exactly this), keep a second model wired in — Opus 5, GPT-5.6 Sol, whatever suits — and know how to fail over in an afternoon. If you’re new to the family history, the glossary entry has it in one page.

The weather around Anthropic isn’t only regulatory, either. In the same week as the launch, Sony Music Publishing and Warner Chappell filed a 48-page complaint naming Anthropic and its founders personally, seeking up to US$150,000 per song. In another courtroom the US Department of Justice argued that training on copyrighted text is fair use. Both things can be true for a long while: the government thinks training is legal, and the publishers are going to keep suing anyway. Neither should change what you do this month. It’s just one more reason not to marry a single vendor.

Do You Actually Need It?

Honestly — maybe not. Anthropic’s line-up below Fable is in good shape. Opus 5 launched on 24 July at US$5/US$25 per million tokens. Sonnet 5 is US$2/US$10, and the price rise to US$3/US$15 that was pencilled in for 1 September was quietly cancelled. Haiku 4.5 sits at US$1/US$5 with a 200K context. The model comparison has the full table if you want it side by side.

A rough rule of thumb

• Drafting, summarising, spreadsheet sanity checks. Sonnet 5 does this well at a fifth of the price. Fable 5.1 is not a better writer of emails.
• Coding and analysis where a person checks the work. Opus 5 is a very good model at half the price. Most of the teams I train live here and don’t realise it.
• Long unsupervised agentic runs, hard debugging, multistep research where you’d rather pay for it to be right first time. This is what 5.1 is for, and at high effort with caching it’s where the 45% figure lives.
• Security or life-sciences work that keeps getting refused. That’s a Mythos question, and the answer is a trusted-access application, not a bigger budget.

If you don’t recognise your work in the third bullet, the upgrade is nice to have rather than necessary. That’s not a criticism of the model. It’s a compliment to the ones beneath it.

The Bottom Line

Fable 5.1 is a good, cheaper-in-practice upgrade for people doing long agentic work, and a mild one for everyone else. Test the API changes, check your plan, keep a fallback.

The thing to watch is the two-door release. If it works, every lab will copy it, and “which model are you on” will start to mean “which version, at which clearance level”. Not a bad question to get comfortable with now, while the answer for most of us is still simply: the one we can buy.

Frequently Asked Questions

What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?

They’re the same model at two safeguard levels. Fable 5.1 is generally available on every platform Anthropic serves. Mythos 5.1 has fewer restrictions and is only available through trusted-access programmes for cybersecurity and life-sciences work. Most people and businesses will only ever use Fable 5.1.

Is Claude Fable 5.1 cheaper than Fable 5?

The list price is unchanged at US$10 per million input tokens and US$50 per million output tokens. What changed is the cost of cache reads, cut to a quarter. Anthropic says that works out at roughly 25% cheaper for typical workloads and up to 45% cheaper for highly agentic work. If your usage has no repeated context, you’ll save little or nothing.

Will my Fable 5 API code work with Fable 5.1?

Mostly, but test first. Forced tool use now returns an error, earlier models can’t read Fable 5.1 thinking blocks, and editing an earlier turn in a conversation invalidates its thinking blocks. Two new beta features arrive alongside: per-message effort settings and turn-scoped system messages.

Is Claude Fable 5.1 included in my Claude subscription?

It depends on the plan. On Claude.ai, Fable 5.1 is included on Max, Team Premium and the premium Enterprise tier. Pro and Team Standard subscribers can use it through usage credits.

Should I upgrade from Claude Opus 5 to Fable 5.1?

Only if your work involves long unsupervised agentic runs, hard debugging or multistep research at high effort. For everyday drafting, analysis and coding where a person checks the output, Opus 5 at US$5/US$25 per million tokens and Sonnet 5 at US$2/US$10 remain very capable and considerably cheaper.