Anthropic

Claude Mythos 5.1

Provider

Anthropic

Status

Restricted

Context

1,000,000 tok

Knowledge

2026-06

Claude Mythos 5.1 is the restricted half of Anthropic’s 1 September 2026 release — the same underlying model as Claude Fable 5.1 with its cybersecurity and biology safeguards relaxed, available only to vetted US organisations. It posts 60.9% on Terminal-Bench 4.0, the top score in Anthropic’s launch table and 5.1 points above Fable 5.1.

It is the most capable model on our board and the one almost nobody can use. We rank it #1 and say so with the same caveat we apply to Mythos 5: a model behind a vetting programme is a capability signal, not a purchase.

Quick specs

What Mythos 5.1 is

Anthropic’s ladder runs Haiku → Sonnet → Opus → Mythos-class. Within Mythos-class it ships two safeguard levels of one model:

The naming carried over from the June release, but the access model changed. Mythos 5 reached its users through Project Glasswing, a US-government cyber-defence and critical-infrastructure programme. Mythos 5.1 runs through two narrower verification programmes instead:

• The Cyber Verification Program — reduced cyber safeguards for defensive security work.
• The Life Sciences Verification Program — reduced biology safeguards for professional research and development, run in partnership with the US government.

Both are limited to US organisations at launch, and Anthropic gave no date for wider access.

The one benchmark

Anthropic published a single figure for Mythos 5.1.

That is the whole disclosure. It is a meaningful result — the top of Anthropic’s agentic-coding table, and a 19-point jump over Mythos 5 — but three things are missing that were present for Mythos 5 in June:

• No ExploitBench or BioMysteryBench figures. Mythos 5’s launch published both (78.0% and 46.1%), and they were the clearest evidence of what removing the safeguards actually buys. Anthropic did not repeat them.
• No SWE-bench figure of any kind, matching the gap on Fable 5.1.
• No independent measurement. Artificial Analysis scores Fable 5.1 at 66 on its Intelligence Index, first of 192 models tracked — but it cannot score a model it cannot access, so there is no independent number for Mythos 5.1 and there will not be one while access stays gated.

Our #1 placement therefore rests on a narrower base than Fable 5.1’s #2: Anthropic’s own Terminal-Bench 4.0 result, plus the structural fact that it is the same base model as Fable 5.1 with fewer safeguards intercepting requests. That is enough to place it above its sibling, and not enough to claim much more.

How to get access

There is no self-serve route. Both programmes are Anthropic-operated verification processes for organisations, not individuals, and both are US-only at launch:

If neither applies to you, Fable 5.1 is the model to use — it is the same base model, and Anthropic reports its retuned safeguards firing around 60% less often per session in Claude Code than Fable 5’s did, with vulnerability discovery now explicitly permitted (exploit development is not).

Mythos-class models remain “Covered Models” carrying mandatory 30-day data retention, with no zero-data-retention option, and outputs carry the invisible numerical watermark Anthropic added to models released after 2 August 2026.

How Claude Mythos 5.1 compares

vs Claude Fable 5.1

Same model, different safeguard levels, and a 5.1-point gap on Terminal-Bench 4.0 (60.9% vs 55.8%) that is best read as the cost of the safeguards rather than a difference in the weights. Fable 5.1 is the one with published pricing, general availability and an independent Intelligence Index score of 66. Unless you are inside one of the two programmes, Fable 5.1 is the practical answer.

vs Claude Mythos 5

Mythos 5 scores 42.0% on the same Terminal-Bench 4.0 test — an 18.9-point gap in under three months. The access story also narrowed: Mythos 5 ran through Project Glasswing, aimed at government cyber-defence and critical infrastructure; 5.1 splits access into a cyber programme and a life-sciences programme, both US-only.

vs Claude Opus 5

Opus 5 is two tiers of access below and 8.6 points behind on Terminal-Bench 4.0 (52.3%), but it has published pricing at $5/$25, a published SWE-bench Pro score of 79.2%, and an independent Intelligence Index score of 63. For anything you actually have to buy and run, that comparison is not close.

Known limitations

You almost certainly cannot use it. Access requires approval through one of two US-only verification programmes. There is no waitlist price, no consumer access and no announced timeline for wider availability.

One published benchmark. Terminal-Bench 4.0 at 60.9% is the entire disclosure in Anthropic’s launch announcement — and unlike Mythos 5, no cyber or bio benchmark accompanied it, which is the exact evidence that would show what relaxing the safeguards buys.

No independent scoring, and there will not be any. Independent evaluators cannot access it.

No published price. Commercial terms are data not available.

Mandatory 30-day data retention. No zero-data-retention option.

Version history

Frequently asked questions

What is Claude Mythos 5.1?

It is Claude Fable 5.1 with its cybersecurity and biology safeguards relaxed — the same underlying model, released on 1 September 2026, restricted to vetted US organisations. It scores 60.9% on Terminal-Bench 4.0 against Fable 5.1’s 55.8%.

Can I use Claude Mythos 5.1?

Only through Anthropic’s Cyber Verification Program (defensive security work) or its Life Sciences Verification Program (professional research and development, run with the US government). Both are for organisations, both are US-only at launch, and Anthropic has given no date for wider access. Everyone else should use Fable 5.1.

How much does Claude Mythos 5.1 cost?

Anthropic did not publish a price. Access is programme-gated rather than sold from a price list. The same base model sells as Fable 5.1 at $10 per million input tokens and $50 per million output.

Why do you rank a model nobody can buy at #1?

Because the board ranks capability and labels access separately — the table marks Mythos 5.1 Restricted for exactly this reason. It is the top of Anthropic’s own launch table and the same base model as the #2 with fewer safeguards in the way. For a buyable recommendation, our best-overall pick is Claude Opus 5; the highest-ranked model you can purchase is Fable 5.1.

Is Mythos 5.1 better than Mythos 5?

On the only benchmark they share, substantially: 60.9% against 42.0% on Terminal-Bench 4.0. Anthropic did not republish the cyber and bio benchmarks that accompanied Mythos 5, so there is no like-for-like comparison on the capabilities the safeguards actually govern.

Last verified 2 September 2026. The Terminal-Bench 4.0 figure is Anthropic-reported and vendor-run, from the 1 September 2026 launch announcement; it is the only benchmark Anthropic published for this model. There is no independent measurement because access is restricted. Pricing is not published. Access terms and availability subject to change.