Overview
Claude Mythos 5.1 is Anthropic's September 1, 2026 successor to Claude Mythos 5. It uses the same underlying model, specifications, and pricing as Claude Fable 5.1, but Anthropic offers it only to vetted organizations whose cybersecurity or life-sciences work needs more permissive safeguards than the generally available model.
The release matters to a narrow audience: approved defensive-security teams, selected life-sciences researchers, and existing Project Glasswing customers. It raises the cyber and biology capability ceiling, improves alignment results over Mythos 5, cuts cache-read pricing, and carries the shared 5.1 API changes. Teams without trusted-access approval should use Fable 5.1 instead.
What's New
Improves Cybersecurity Capability
Anthropic describes Mythos 5.1 as the strongest cybersecurity model it has released. With cybersecurity safeguards disabled for evaluation, Mythos 5.1 reaches 60.9% on Terminal-Bench 4.0; the same launch table reports 55.8% for Fable 5.1 and 42.0% for Fable 5. The company still places Mythos 5.1 within the lower category of cyber risk in its Frontier Compliance Framework and restricts deployment through trusted-access programs.
Claude Security, Anthropic's codebase vulnerability scanner and patch-suggestion product, now runs on Mythos 5.1. That product use does not make the raw Mythos model generally available.
Expands Life-Sciences Research Results
In Anthropic's reported external validation, Mythos 5.1 designed protein binders with nearly a 50% hit rate across 12 targets, versus the 10-15% rate Anthropic describes as typical. On three targets, binding affinities were ten times higher than the best submissions to Adaptyv Bio competitions.
The model also optimized seven open-source protein and genomics models with identical outputs, producing inference speedups of up to 2.5x. Anthropic estimates 30-60% lower GPU cost for selected genome-wide analyses, although these results are workload-specific and the optimizations had not yet been open-sourced at launch.
Improves Alignment While Keeping Trusted Access
Anthropic's automated behavioral audit found Mythos 5.1 better aligned than Mythos 5 across most measured areas. It was less likely to access resources outside its test environment, use motivated reasoning, ignore explicit constraints, or attempt and succeed at reward hacking. The company also reports its strongest result to date on an external prompt-injection benchmark.
These improvements do not remove the access gate. Anthropic says Mythos 5.1 has greater chemical and biological capabilities than Mythos 5 and retains the same safeguards that restrict access to research biology capability.
Cuts Cache-Read Pricing
Like Fable 5.1, Mythos 5.1 keeps base input and output prices at $10 and $50 per million tokens while reducing cache reads to $0.25 per million tokens. Long research and agent sessions that repeatedly read a stable cached prefix pay one quarter of the Mythos 5 cache-read rate.
Adds Shared 5.1 Agent Controls and Provenance
Approved API customers receive the same 5.1 controls as Fable 5.1, but per-message effort, turn-scoped system messages, and the thinking.display: "updates" option are beta features. Progress-update blocks themselves are not new to Fable 5.1; the new 5.1 capability is the beta option that exposes those updates as readable text.
Text generated by Mythos 5.1 carries Anthropic's statistical watermark. On the Claude API, supported image and video files Claude produces carry signed C2PA Content Credentials when retrieved through the Files API.
Performance & Research Evidence
| Area | Mythos 5.1 result | Previous reference | Decision value |
|---|---|---|---|
| Terminal-Bench 4.0 | 60.9% | Fable 5: 42.0% | Re-test approved cyber agent workflows |
| Protein-binder hit rate | Nearly 50% across 12 targets | Typical rate cited by Anthropic: 10-15% | Potentially fewer failed design candidates |
| Binding affinity | 10x higher on three targets | Best Adaptyv competition designs | Candidate quality for selected targets |
| Biology model inference | Up to 2.5x faster | Original open-source implementations | Faster repeated research analysis |
| Estimated GPU cost | 30-60% lower on selected analyses | Original cloud-list-price workload | Lower cost for repeated genome-wide runs |
These are Anthropic-published results and external validations selected for the launch announcement. They do not replace independent reproduction, laboratory validation, security controls, or organization-specific evaluations.
Compared With Previous Version
| Area | Mythos 5 | Mythos 5.1 |
|---|---|---|
| Model ID | claude-mythos-5 |
claude-mythos-5-1 |
| Underlying public counterpart | Fable 5 | Fable 5.1 |
| Terminal-Bench 4.0 | Not separately reported in the 5.1 launch table | 60.9% |
| Life-sciences access | Limited Glasswing access; biology program planned | Invite-only Life Sciences Verification Program launched with initial participants |
| Cyber access | Vetted defensive-security organizations | Available to approved Project Glasswing cyberdefenders; the Cyber Verification Program is planned to add Mythos access in the near future, and Claude Security already runs on Mythos 5.1 |
| Base API price | $10 input / $50 output per MTok | Unchanged |
| Cache reads | $1 per MTok | $0.25 per MTok |
| Forced tool choice | Supported | any and named tool choices return HTTP 400 |
| Thinking prefix check | Previous Mythos behavior | Does not run Fable 5.1's conversation-prefix check |
| Provenance | No 5.1 provenance behavior | Statistical text watermark and C2PA for supported files |
Migration Guide
Confirm 5.1 Access Before Cutover
Changing the model name does not grant access. Existing customers should confirm Mythos 5.1 entitlement with their Anthropic, AWS, or Google Cloud account team and verify which cloud platform and trusted-access program covers the workload.
Update the Model ID
Replace claude-mythos-5 with claude-mythos-5-1. The Claude API and Google Cloud use claude-mythos-5-1, while Amazon Bedrock uses anthropic.claude-mythos-5-1. Re-run end-to-end evaluations because model behavior and tool orchestration differ from Mythos 5 even when request schemas remain valid.
Remove Forced Tool Choice
Mythos 5.1 shares Fable 5.1's forced-tool breaking change. Requests using tool_choice type any or a named tool return a 400 invalid_request_error. Use automatic tool choice, explicit instructions, and strict tool schemas or structured outputs instead.
Preserve Thinking Blocks Across Model Switches
Earlier Claude models cannot read 5.1 thinking blocks, so routers that fall back from Mythos 5.1 can lose those blocks for turns handled by the older model. Unlike Fable 5.1, Mythos 5.1 does not run the conversation-prefix check that rejects or drops thinking blocks after earlier history changes. Even so, passing blocks back unchanged and testing routing transitions is the safer migration path.
Re-Tune Effort and Parallel Tool Use
Adaptive thinking remains always on, with high as the default API effort. Test lower effort for routine work and higher effort for difficult research steps. Monitor long agent loops for fewer parallel tool calls and make independent work explicit when round trips and token cost matter.
Availability & Access
Mythos 5.1 is invite only. Anthropic's launch materials describe two routes:
- The Life Sciences Verification Program is an invite-only beta for advanced researchers using reduced biology safeguards. Anthropic enrolled its first participants with the US government and plans broader expansion.
- The Cyber Verification Program provides reduced cyber safeguards for defensive work and is expected to include Mythos-class access in the near future.
At launch, Anthropic said it could offer Mythos 5.1 only to a set of US organizations. Applicants outside that set should not assume approval or a delivery date. Claude Security access is separate from direct API entitlement.
Mythos 5.1 requires 30-day data retention for safety monitoring by default. The official model page lists Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry availability for approved customers.
Pricing & Plans
| Item | Price or requirement |
|---|---|
| Input | $10 per million tokens |
| Output | $50 per million tokens |
| Five-minute cache write | $12.50 per million tokens |
| One-hour cache write | $20 per million tokens |
| Cache read | $0.25 per million tokens |
| Batch API | 50% discount on input and output |
| Access | Invitation and trusted-program approval required |
Published token rates match Fable 5.1, but Mythos cannot be purchased on demand. Approval, program conditions, cloud availability, and retention requirements are part of the actual adoption cost.
Who Should Upgrade / Who Should Wait
Upgrade only if your organization has confirmed Mythos 5.1 access and specifically needs its more permissive cybersecurity or life-sciences safeguards. Existing Mythos 5 access alone does not grant Mythos 5.1 entitlement, and general long-horizon capability alone is not a reason to choose Mythos because Fable 5.1 uses the same underlying model.
Wait if you lack trusted-access approval, operate outside the currently supported organization set, require default zero data retention, or can complete the work under Fable 5.1's generally available safeguards.
Known Limitations
- Access remains limited and is not guaranteed by paying the published token rate.
- The Cyber Verification Program had not yet added Mythos access for general applicants at launch.
- Forced named or
anytool choice is no longer supported. - Raw 5.1 thinking blocks are not readable by earlier models during a fallback or downgrade.
- The model requires 30-day retention by default.
- Cyber and biology results require independent security, scientific, and laboratory validation.
- Anthropic reports less alignment-evaluation coverage for very long-context, multi-agent, and impossible-task settings.
FAQ
Is Mythos 5.1 a more powerful model than Fable 5.1?
No. Anthropic describes them as the same underlying model. The difference is the safeguard and access policy: Mythos provides more permissive controls for vetted cybersecurity and life-sciences work.
Can any enterprise buy Mythos 5.1 at $10/$50 per million tokens?
No. Those are the starting token rates after access is approved. Mythos 5.1 remains invite only and was limited to a set of US organizations at launch.
Does Mythos 5.1 have the same migration risks as Fable 5.1?
Most shared API changes apply, including the forced-tool incompatibility and model-switch handling for thinking blocks. Mythos 5.1 does not run Fable 5.1's conversation-prefix check, which is the main exception documented by Anthropic.
