Claude Fable 5.1 vs Mythos 5.1: same model, different safeguards

Claude Fable 5.1 vs Mythos 5.1: one model, two safeguard tiers

The single most confusing thing about this release is that Fable 5.1 and Mythos 5.1 are not two models. Anthropic states plainly that they are the same model with different levels of safeguards. Fable 5.1 is the generally available build with production safeguards on. Mythos 5.1 is the same weights with the cybersecurity and life-sciences restrictions loosened, released only to vetted organisations.

So the split is not a capability tier. It is a permission tier.

Note the last row. Anthropic's own pricing documentation lists Mythos 5.1 at exactly the same rates as Fable 5.1, flagged as limited availability. You do not pay more for Mythos. You get vetted for it, or you do not get it.

Where Fable and Mythos sit in the Claude model ladder

Anthropic's naming now runs across five names for four price points. Haiku, Sonnet and Opus are the small, medium and large models; Fable and Mythos sit above Opus at the top tier, at double Opus rates.

One quiet change worth catching if you budgeted for it: Sonnet 5's $2 / $10 rate launched as introductory pricing due to expire on 31 August 2026. Anthropic has now made it the standard price, and the scheduled increase to $3 / $15 will not happen. If your routing logic was written to fall back off Sonnet in September, you can delete that branch.

What Claude Fable 5.1 costs after the cache-read price cut

Base rates did not move. Fable 5.1 is still $10 per million input tokens and $50 per million output tokens, the same as Fable 5. The entire price story is one line item: cache reads.

Every other Claude model prices a cache hit at 0.1x base input. Fable 5.1 and Mythos 5.1 price it at 0.025x. That is the whole mechanism behind the headline: cache reads dropped 75%, from $1 to $0.25 per million tokens, and Anthropic estimates that produces roughly 25% savings on typical workloads and up to 45% on highly agentic ones.

A second-order effect that the launch post does not spell out: the multiplier change also moves the point at which caching pays for itself. On a standard 0.1x model, a 5-minute cache write (1.25x) breaks even after one read and a 1-hour write (2x) after two. On Fable 5.1, a read saves 0.975x instead of 0.9x, so the 5-minute write clears its cost with room to spare on the first read, and the 1-hour write clears on the second. Short-lived caches that were marginal on Fable 5 are now worth taking.

The trap is symmetrical. The discount only touches tokens that come back as cache hits. If your bill is dominated by output tokens, or by fresh input the cache never sees, you will see close to none of the 25%.

Is Fable 5.1 cheaper than Opus 5 for agent workloads?

Per cached token, yes: $0.25 against Opus 5's $0.50. Per request, almost always no. It depends entirely on the shape of your bill, so it is worth doing the arithmetic on a realistic loop rather than trusting the percentage.

Take an agentic coding session with a stable 100k-token prefix (system prompt, tool definitions, repository context) cached on the 1-hour window, 40 turns in that hour, 2k of fresh input and 1.5k of output per turn.

Three things fall out of that table. The Fable 5 to Fable 5.1 saving is 31%, close to Anthropic's estimate. Fable 5.1 is still 1.39x the cost of Opus 5, down from 2.0x on Fable 5 — the gap narrowed sharply but did not close. And the cache-read line, previously the largest single cost in the session, is now the smallest.

That last point is the one that should change your architecture. On Fable 5, stuffing more context into the cached prefix made the loop more expensive with every turn, which pushed teams toward aggressive context trimming. On Fable 5.1, a larger cached prefix is close to free to re-read, and output length becomes the dominant cost lever. If you have been trimming context to control spend, you are now optimising the wrong line.

One caveat when you compare against historical bills: Claude 4.7 and later use a tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. A flat per-million comparison against an older model understates what you will actually pay. If you are re-costing an agent pipeline around these rates, that modelling work is exactly the kind of thing our AI automation and workflow team does before committing a client to a model.

Fable 5.1 benchmark results, and the caveat in the footnotes

Anthropic's published comparison against Fable 5, Opus 5 and OpenAI's GPT-5.6 Sol:

The Terminal-Bench 4.0 row is the interesting one, because it contains both names. Fable 5.1 scores 55.8% and Mythos 5.1 scores 60.9% on identical weights. The five-point gap is not capability. It is the cost of the safeguards firing on benchmark tasks that touch cybersecurity, where Fable was scored on a redirected answer.

Anthropic is unusually explicit about this in its footnotes: Fable 5.1 was evaluated with production safeguards enabled, scored zero on the OSWorld 2.0 tasks where safeguards intervened, and had intervened cyber tasks completed by Claude Opus 4.8 and intervened biology tasks by Claude Opus 5. Read the practical implication rather than the marketing one — these are the scores of the system you actually get, safeguards included, which is more honest than a benchmark run with the guardrails switched off. It also means the published Fable numbers understate the raw model and overstate what a redirect-heavy workload will feel like.

Terminal-Bench-Science is worth a second look too. Anthropic notes a standard error of ±3.5–4.5 points per model, so read a 52.6% against 29.0% as a real gap and any two-point difference as noise.

Why some Fable 5.1 requests are still answered by an Opus model

This is the operational detail most launch coverage skipped, and the one most likely to surprise you in production. Fable's safeguards do not only refuse. For several categories they silently route the request to a different, smaller model.

What changed with 5.1:

• Cyber safeguards now fire around 60% less often per Claude Code session than the safeguards that launched with Fable 5.
• Fable 5.1 can now be used to identify software vulnerabilities — defensive work — though not to develop exploits for them.
• Biology safeguards fire 85% less often on benign elementary biology and medical questions.

What still gets redirected to Opus models:

• Penetration testing
• Exploit generation
• Binary-based vulnerability scanning
• Life-sciences research and development queries

If you are building an agent in one of these domains, design for it. A pipeline that assumes every response came from the model you named will produce inconsistent output quality, inconsistent token accounting and inconsistent latency across a run, and the inconsistency will correlate with topic rather than with load. Log which model answered, and treat a redirect as a routing event your system knows about rather than an anomaly your users discover.

Anthropic also says two external organisations plus Gray Swan stress-tested these safeguards and found no critical-severity jailbreak, so the redirect behaviour is not something to plan around evading.

How to get access to Claude Mythos 5.1

Two programs, both gated:

• Cyber Verification Program (CVP) — currently provides reduced cyber safeguards on Opus- and Sonnet-class models for defensive security work, with Mythos-class access being added. Applications go through the Anthropic programs portal.
• Life Sciences Verification Program (LSVP) — built in partnership with the US government, first participants already enrolled, broader life-sciences enrolment expected.

Availability is currently limited to a set of US organisations, with Anthropic saying it is coordinating with the US government to extend access to more domestic and international partners. Anthropic's Claude Security product, which scans codebases and suggests patches for human review, now runs on Mythos 5.1 — which is a more realistic route to Mythos-grade security analysis for most teams than applying for direct model access.

The invisible watermark on Claude's text, and what it means for content teams

Fable 5.1 ships with text watermarking, and this is where a lot of content and SEO teams are drawing the wrong conclusion.

The facts, from Anthropic's own FAQ: the watermark exists to comply with the EU AI Act's Code of Practice on Transparency of AI-Generated Content, which Anthropic signed in July 2026 along with around 190 other signatories. It applies to models released after 2 August 2026. It works by seeding token sampling from a key rather than an arbitrary random source, so the output stays statistically random but carries a pattern anyone holding the key can test for. It is invisible to readers and to ordinary AI-detection tools, and it encodes nothing about the user, their organisation or their conversations.

Anthropic is rolling out a detection API in private preview, currently limited to eligible organisations — regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil society groups — plus enterprises with their own compliance obligations under the Act.

Two implications worth separating. First, this is a compliance mark, not a quality signal: it says a model was involved in producing text, not that the text is bad, and search engines have no access to the key. Second, the honest response for a publisher is the one that was already correct before watermarking existed — disclose how content is produced, and put enough human expertise, verification and original judgement into a piece that the question of who typed the first draft stops being the interesting one.

The anti-distillation change that can break existing integrations

Buried in the safety section is a change with direct engineering consequences. To make distillation attacks harder, API accounts created from the Fable 5.1 launch onward can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of Claude's prior thinking.

Existing accounts are not affected yet, but Anthropic says the restriction will apply to all users with future model releases, and that a small number of custom integrations will break when it does.

If your system rewrites assistant turns before replaying them — trimming tool outputs, compacting history, injecting corrected reasoning — check whether it depends on retaining thinking blocks through that edit. It is a cheap thing to audit now and an expensive thing to discover on a launch day six months from now.

Who should switch to Fable 5.1, and who should not

Anthropic notes one behavioural default that quietly affects both cost and quality: Fable 5.1 runs at High effort in Claude Code, and Medium in Claude Cowork and on Claude.ai. If you benchmark it in the app and then deploy through the API without setting effort explicitly, you are not comparing the same thing.

The broader pattern this release confirms is that model choice is no longer a single decision. It is a routing problem across at least four price points, several effort levels, and a safeguard layer that can hand your request to a different model mid-run. If you would rather have that architecture designed and costed once, properly, than rebuilt every time a version number ticks over, tell us what you are building and we will scope it.