Claude Fable vs Claude Mythos: Same Weights, Same Price, and a Choice You May Not Get to Make

Claude Fable 5.1

The two builds, described as Anthropic describes them

Claude Fable 5.1 shipped on 1 September 2026 as the successor to Claude Fable 5, in the tier Anthropic positions above the Opus class for demanding reasoning and long-horizon agentic work. It is generally available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, with code>claude-fable-5-1/code> as the model ID and a retirement commitment of not sooner than 1 September 2027.

Claude Mythos 5.1 shipped the same day on the same weights, with the cyber and biology safeguards tuned down for vetted defenders. It is not in the general catalogue. Access runs through the Cyber Verification Program — rebuilt on 6 October into Defense, Red Team and Specialized tiers — or the separate, invite-only Life Sciences Verification Program, and at launch it was limited to US organisations with mandatory 30-day data retention.

The generation before them follows the identical pattern: Claude Fable 5 and Claude Mythos 5, both 9 June 2026, both suspended worldwide two days later by a US Commerce Department export-control directive, both restored for a limited set of vetted organisations around the start of July.

What is genuinely different, row by row

• Price — identical. $10 input, $50 output, $0.25 cache read, half on batch, and a 1.1× multiplier for US-only inference on the Mythos track that Fable does not carry.

• Safeguards — the whole difference. Fable runs the full cyber and biology classifier stack; a declined request returns HTTP 200 with a code>refusal/code> stop reason you have to check before reading content. Mythos reduces those interventions, which is why the review exists.

• Access — self-serve versus an application. Fable is a purchase order. Mythos is a verification review, a tier assignment, and roughly five business days of provisioning after approval.

• Data retention — not a difference, and this surprises people. Fable 5.1 is also unavailable under zero data retention without express authorisation from Anthropic; an organisation whose retention configuration does not meet the requirement gets a 400. Both builds are 30-day models.

• Thinking blocks — a real API-level split. Only Claude Fable 5.1 and Claude Mythos 5.1 can read Claude Fable 5.1's thinking blocks. The two are also treated differently on the history-editing check that preserved thinking introduced: Fable 5.1 enforces it on replayed thinking blocks, Mythos 5.1 does not run that check at all. If you are building an append-only harness, that difference is not academic.

• Request surface — otherwise the same. Forced tool use returns a 400 on both, assistant prefill is gone on both, thinking cannot be disabled on either, and effort runs low through max with code>high/code> as the default.

What the safeguard tier buys, in one experiment

Anthropic published the closest thing to a controlled comparison of the two access levels, and it used Claude Opus 5.5 rather than either Fable or Mythos to do it. The company ran ten multi-stage cyber-operation challenges, five attempts each, once per access tier.

• With no program access — every task blocked at the first prompt. Zero attempts started.

• Inside Defense Access — 46 of the 50 trials blocked somewhere in the run.

• Inside Red Team Access — 34 of 50 completed, effectively matching the 67.6% success rate the same model achieves with no safeguards applied at all.

Read that as a description of what the gate is buying. The capability is the same model in all three arms; what changed is how often an intervention stopped it. Defense Access is not a weaker Mythos — it is a model that gets in the way most of the time. Red Team Access is close to the unfiltered number.

There is a cost side that Anthropic also put on the record. Project Glasswing partners reported at least 129,000 verified software vulnerabilities between April and July 2026, its own open-source scanning added 5,500 more between April and October, and more than 33,000 of the combined total were rated critical or high severity. Those are the company's figures, drawn from partial partner reporting, and Anthropic says the real number is likely at least five times higher. Whatever you think of the estimate, it is the argument for keeping a reduced-safeguard build in the hands of defenders.

Where the benchmark comparison goes wrong

The temptation is to line up Claude Fable 5.1 and Claude Mythos 5.1 on one table. Anthropic makes that impossible on purpose: the published comparison table pairs Fable 5.1 with Fable 5, and the company says those deltas carry across to the Mythos track, where the headline agentic figure is 60.9% on Terminal-Bench 4.0 against 42.0% for the previous generation.

What you cannot do is compare that 60.9 to Claude Mythos 5's better-known 88.0 — because 88.0 is Terminal-Bench 2.1, a different test on a different harness. The same trap catches OSWorld: Anthropic reports 85.0% for Mythos 5 on OSWorld-Verified, while other vendors quote OSWorld 2.0 scores in the sixties. Two numbers, two tests, no shared meaning.

The one independent reading comes from Artificial Analysis, which scores Claude Fable 5.1 at 53 on its Intelligence Index — the same model, in two evaluation modes, and both around that figure. Mythos 5.1 does appear on the index; Mythos 5 does not. On the leaderboard's own consolidated view, Claude Opus 5.5 sits above Fable at 58, which is worth holding next to the price: the cheaper Opus build scores higher on that particular ruler.

The decision, stated plainly

If both builds were on the same shelf at the same price, almost nobody would take Fable. Mythos scores higher on the agentic track, has fewer interventions in the way, and costs the same. That is exactly why it is not on the shelf.

So the honest decision rule is inverted from a normal comparison:

• You are a security organisation doing defensive work, incident response, malware analysis or authorised red-teaming — apply. Defense Access is open to individuals on a paid plan, which means a single researcher with a track record can get in. Start now rather than when you need it; the review clock, not the model, is the long pole.

• You are everyone else — Fable 5.1 is the same model with the classifiers left on, and it is the one you can actually buy. For agentic coding, long-document work and multi-step research the safeguards rarely fire, and the cost structure is identical, so the only thing you are giving up is the small set of requests a classifier would have declined.

• You need the capability without the retention terms — you have a problem neither build solves. The 30-day retention requirement applies to both, and zero-data-retention organisations get a 400 on Fable 5.1 just as they would on the Mythos side.

Claude Fable 5.1 is in the OrcaRouter catalogue as code>anthropic/claude-fable-5.1/code> at Anthropic's list price with 0% markup — $10 and $50, with the $0.25 cache read — and Claude Fable 5 is there beside it. Running both through one key is how you measure the 5.1 cost structure on your own workload without a second contract, and automatic failover covers the case where a provider has a bad afternoon. Claude Mythos 5.1 is not in our catalogue and will not be: the verification programme is the gate, and a router does not get to route around a review.

What would make this a real choice

One publication would do it: an independent evaluation that runs Claude Fable 5.1 and Claude Mythos 5.1 through the same harness and reports both, so the safeguard cost stops being a vendor figure. Until then the comparison rests on one sentence in Anthropic's own documentation and a run of CyScenarioBench on a third model.

The thing to stop doing is treating Mythos as the better model you cannot have. It is the same model with a narrower door, and for most workloads the door is the only part you would ever notice.

Sourcing note: the statement that Claude Mythos 5.1 shares Claude Fable 5.1's specifications and pricing, the thinking-block rule, the retention requirement and the request-surface restrictions are from Anthropic's Claude Platform documentation, read on 26 October 2026. The CyScenarioBench trial counts, the 67.6% unsafeguarded figure, the Terminal-Bench 4.0 and OSWorld-Verified scores, the Fable 5.1 index deltas and the Project Glasswing vulnerability counts are Anthropic's published material from the same period and are vendor-reported. The Intelligence Index figure of 53 for Claude Fable 5.1 and 58 for Claude Opus 5.5 are Artificial Analysis's, read on the same date. Prices and route measurements are OrcaRouter's catalogue and our own seven-day traffic.