Citation and evidence

Claude Mythos 5.2

Report a problem with this article

Edit article
More

Cite this article

Claude Mythos 5.2 is the verified-access configuration of an Anthropic language model released on November 9, 2026, alongside Claude Fable 5.2. The two share an underlying model, but Mythos permits a broader set of cybersecurity and biology tasks through controlled access programs. It succeeds Claude Mythos 5.1.[1][2]

Direct access depends on approval, not simply on selecting a model name or buying a subscription. The applicable programs set the permitted work and access conditions. Anthropic also uses Mythos in its Claude Security product, which provides a separate, constrained route to model-powered code scanning.[1][3][5]

Relationship to Fable

Fable is the configuration offered for general use, with additional safeguards for sensitive cyber and biology work. Mythos is offered to vetted users whose work qualifies under Anthropic's verification programs. This is a deployment and safeguards distinction, rather than a difference in the documented context limit, output limit, or base API price.[1][2]

The shared model does not make the two products' measured results interchangeable. Anthropic labels its public benchmark table as Fable results unless a row explicitly identifies Mythos. Differences on affected tasks can reflect Fable's classifiers and fallback behavior, rather than a separate set of underlying model weights.[1]

Specifications

PropertyDocumented specification
Release dateNovember 9, 2026
Claude API model IDclaude-mythos-5-2
Input and outputText and images in; text out
Context window1 million tokens
Maximum output128,000 tokens
ThinkingAdaptive, always on
API default efforthigh
Reliable knowledge cutoffJune 2026
StatusActive; verification required
Retirement commitmentNot sooner than November 9, 2027

Expanded article table

Anthropic lists the Claude API, Amazon Bedrock's bedrock-runtime endpoint, Anthropic-managed and Google-managed Claude offerings on Google Cloud, and Microsoft Foundry. Bedrock uses identifiers such as global.anthropic.claude-mythos-5-2. A platform listing does not by itself grant model access.[2]

Access programs

Cyber Verification Program

The Cyber Verification Program distinguishes three tiers. Its detailed eligibility documentation, checked November 17, 2026, explicitly includes approved individuals in Defense Access; Red Team and Specialized Access are organization-only.[3]

TierGeneral scopeWho can apply
Defense AccessDefensive operations, incident response, analysis, and vulnerability validationIndividuals and organizations
Red Team AccessAuthorized penetration testing and red-team workOrganizations
Specialized AccessApproved high-risk testing of systemically important infrastructureOrganizations

Expanded article table

Individual Defense applications can use eligible paid plans or the API; the Free plan is excluded. Organization-only tiers require Team, Enterprise, or API access. Approval applies to the approved account or organization, and does not automatically transfer to another one.[3]

The program generally covers internal work. Client services and products require separate external-use approval. Verification adjusts safeguards within the approved scope; it does not suspend the Usage Policy or guarantee that every allowed request avoids a classifier block.[3]

Life Sciences Verification Program

The Life Sciences Verification Program is a beta program offering more permissive biology safeguards for qualified professional work. It currently accepts verified organizations on Team, Enterprise, and the Claude Platform API. Free, Pro, Max, and Ultra plans are not eligible for this program.[4]

An organization's administrator registers interest, after which Anthropic assesses eligibility. Completing the interest form does not guarantee participation. These conditions differ from the Cyber Verification Program's individual Defense Access, so eligibility for one should not be assumed to establish eligibility for the other.[3][4]

Claude Security access

Claude Security is an Enterprise public-beta application that scans repositories, validates findings, and proposes patches for human review. Its documentation says that Mythos-powered scans in the Claude.ai Security app do not require separate Cyber Verification Program access. This product access does not establish unrestricted direct use of the Mythos API.[5]

The Claude Security plugin for Claude Code is different: it uses the models available in the user's Code account. Installing that plugin does not independently grant Mythos access.[5]

API pricing

These USD list prices were checked on November 17, 2026. Figures are per million tokens; provider-specific terms, tool use, and geographic modifiers can change the total charge.[6]

Processing modeInputOutput
Standard$8$40
Message Batches$4$20
Fast mode$16$80

Expanded article table

Standard five-minute prompt-cache writes cost $10 per million tokens, one-hour writes cost $16, and cache reads or refreshes cost $0.20. Mythos 5.2 and Fable 5.2 have the same base rates.[6]

Fast mode uses a faster inference configuration of the same model, with advertised output throughput up to 2.5 times standard mode. The throughput claim does not promise that a complete tool-using task finishes 2.5 times faster. Access remains subject to Mythos verification, and fast-mode availability varies by platform. It is not supported by the Batch API.[1][7]

Reported capabilities

Anthropic publishes separate Mythos results for some cybersecurity and biology evaluations. The following table includes only rows explicitly identified as Mythos 5.2, rather than relabeling Fable's general coding and reasoning scores.[1]

EvaluationMythos 5.2 reported result
CyScenarioBench76.3%
CVE Rediscovery89.6%
De novo binder design, 24-hour budget85.1%
Terminal-Bench-Science 0.175.3%

Expanded article table

These are reported evaluation scores, not a universal success rate for security work or biological research.[1][8]

Security evaluation conditions

CVE Rediscovery used 1,829 selected October 2026 advisories. The model received a vulnerable repository snapshot and a list of files changed by the fix, without internet or commit history. A judge checked whether a report matched an advisory's root cause; an exploit was not required. The system card acknowledges possible training contamination and sensitivity to how many reports are scored.[8]

CyScenarioBench used a private ten-challenge subset and a prerelease snapshot. Its cyber mitigations were disabled. The launch table warns that different labs' CyScenarioBench scores use different tasks, so cross-lab percentages should not be treated as a common standardized test.[1][8]

Scientific evaluation conditions

The binder-design evaluation was a computational design campaign with tools, a GPU, and a 24-hour budget per target; its score does not establish a corresponding experimental or clinical success rate. Terminal-Bench-Science used a no-internet agent setup, with ten trials per task. Anthropic reports 75.3% for Mythos without the safeguards used in Fable's 74.6% run.[8]

Protein-structure demonstration

The launch described Mythos 5.2 using existing folding models and known symmetries to predict adeno-associated-virus capsid structures. Anthropic reported meeting its accuracy criterion on 28 of 34 attempted capsids. Its footnote defines an accurate result as either tested folding model producing a structure within 3 angstroms of the experimental reference under the specified comparison.[9]

The work concerns protein folding and computational structure prediction. It should not be interpreted as demonstrated effectiveness or safety of a gene therapy. The launch also acknowledges earlier research on modeling large symmetric protein assemblies.[9]

API integration

Mythos 5.2 shares Fable 5.2's principal API changes. Integrations need to preserve bound thinking blocks, use supported tool declarations, and check advisor-model compatibility. An executor running either 5.2 configuration requires a Fable 5.2 or Mythos 5.2 advisor. Forced tool_choice values any and tool are unsupported.[10]

When migrating from Mythos 5.1, developers should confirm approval for 5.2 before changing model to claude-mythos-5-2. They should also test conversation replay and model routing. Preserved thinking from 5.2 is not readable by older models; routing back to 5.1 can therefore lose usable reasoning for the affected turn.[11]

The conversation binding covers the system prompt, tool definitions, and preceding messages. Editing those earlier elements can invalidate later thinking blocks and result in a 400 error. Passing assistant blocks back unchanged and keeping history append-only avoids that class of mismatch; account and platform compatibility also matter.[12]

The same effort and prompting considerations described for Fable apply to Mythos. Anthropic recommends workload evaluation rather than assuming that max always supplies the best cost-quality tradeoff. Its prompting guide describes premature stopping, sparse progress messages, and unnecessary delegation as behaviors that can still occur on demanding agent tasks.[13]

Data retention and monitoring

Mythos 5.2 is a Covered Model. Anthropic's default policy retains prompts and completions for at least 30 days for safety monitoring, with exceptions to deletion for investigations or legal requirements. These conditions follow the model across supported surfaces.[14]

The API documentation states that zero data retention is unavailable for Mythos 5.2 unless Anthropic expressly authorizes it. A feature's general ZDR eligibility does not override that model-specific condition.[15]

Enterprise Frontier Safeguards is designed to run automated monitoring with retained activity data in customer-controlled cloud infrastructure. Anthropic describes eligibility-based transition arrangements for some customers with existing ZDR terms. Neither verification-program membership nor an Enterprise subscription should be taken to grant an unapproved retention exception.[1][16]

Continue reading

More to read

Topics:AI ModelsAI in CybersecurityAnthropicLarge Language ModelsReasoning Models

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Anthropic. "Claude Mythos 5.2". Model page. Accessed November 17, 2026.
  2. ^1 ^2 ^3Anthropic. "Claude Mythos 5.2: Overview". Claude Platform Docs. Accessed November 17, 2026.
  3. ^1 ^2 ^3 ^4 ^5Anthropic. "Cyber Verification Program". Claude Help Center. Accessed November 17, 2026.
  4. ^1 ^2Anthropic. "Life Sciences Verification Program". Accessed November 17, 2026.
  5. ^1 ^2 ^3Anthropic. "Claude Security". Product documentation and FAQ. Accessed November 17, 2026.
  6. ^1 ^2Anthropic. "Pricing". Claude Platform Docs. Accessed November 17, 2026.
  7. ^Anthropic. "Fast mode". Claude Platform Docs. Accessed November 17, 2026.
  8. ^1 ^2 ^3 ^4Anthropic. "Claude Fable 5.2 & Claude Mythos 5.2 System Card". November 9, 2026. Sections 3.3.2, 8.6, 8.18.4, and 8.19.
  9. ^1 ^2Anthropic. "Claude Fable 5.2 and Mythos 5.2". November 9, 2026. Protein-folding section and footnote 3.
  10. ^Anthropic. "What's new in Claude Fable 5.2". Shared Fable 5.2 and Mythos 5.2 specifications and API changes. Accessed November 17, 2026.
  11. ^Anthropic. "Claude Fable 5.2 and Claude Mythos 5.2 migration guide". Accessed November 17, 2026.
  12. ^Anthropic. "Preserved thinking". Claude Platform Docs. Accessed November 17, 2026.
  13. ^Anthropic. "Prompting Claude Fable 5.2". Applies to both Fable 5.2 and Mythos 5.2. Accessed November 17, 2026.
  14. ^Anthropic. "Covered Models". Claude Help Center. Accessed November 17, 2026.
  15. ^Anthropic. "API and data retention". Claude Platform Docs. Accessed November 17, 2026.
  16. ^Anthropic. "Developing Enterprise Frontier Safeguards with our customers". September 1, 2026.

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 1,681 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "Claude Mythos 5.2." aiwiki.ai, updated 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/claude_mythos_5_2

Edit article