Claude Mythos is Anthropic's restricted-access line of Claude models, offered only to vetted organisations. This profile covers Claude Mythos Preview (announced 7 April 2026 and now deprecated), Claude Mythos 5, Claude Mythos 5.1 and Claude Mythos 5.2.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Anthropic's public documentation as of November 2026. Model details change, so confirm against Anthropic's current documentation.
Key takeaways
- Access is restricted on every Mythos model. Anthropic's model pages list Mythos 5, 5.1 and 5.2 as "Active (verification required)", available only to organisations verified through Anthropic's verification programmes, such as the Cyber Verification Program.
- Each Mythos release pairs with a Fable release. Anthropic describes Mythos 5 as the same model as Claude Fable 5, and Mythos 5.1 and 5.2 as carrying the same specifications and pricing as Fable 5.1 and 5.2, with more permissive safeguards in cyber and biology.
- Mythos 5.2 costs $8 input and $40 output per million tokens. Mythos 5 and 5.1 are $10 and $50; Batch API requests are 50% of standard rates and fast mode for Mythos 5.2 is $16 and $80.
- Context is 1M tokens with 128K output. Mythos 5, 5.1 and 5.2 take text and image input, return text and think adaptively at all times.
- Mythos Preview is deprecated. Anthropic deprecated claude-mythos-preview on 9 June 2026 with retirement "to be announced"; the numbered models carry "not sooner than" retirement dates in 2027.
What is Claude Mythos?
Claude Mythos is a line of proprietary large language models from Anthropic, first released as Claude Mythos Preview in April 2026 under Project Glasswing and now at version 5.2, released on 9 November 2026. The Mythos 5 model page subtitle reads "Most capable model for cybersecurity and biology research" (Anthropic's wording), and the page says it is offered for defensive cybersecurity work.
The Claude Fable 5.2 and Mythos 5.2 system card says the two are configurations of the same new model: Fable 5.2 is the general-use release with added safeguards that block certain high-risk biology and cybersecurity tasks, while Mythos 5.2 has more permissive safeguards in those domains. For the generally available counterpart see the Claude Fable 5.2 profile, and for every Claude model see Anthropic Models.
Anthropic's Glasswing page states that it formed Project Glasswing because of Mythos Preview's ability to find and exploit software vulnerabilities, and that the aim is to give defenders an advantage.
Key facts at a glance
As of 2026-11-15.
| Field | Value | Source |
|---|---|---|
| Provider | Anthropic | Model pages; Glasswing page |
| Claude API ID (Mythos Preview) | claude-mythos-preview | Model pages; Glasswing page |
| Claude API ID (Mythos 5) | claude-mythos-5 | Model pages; Glasswing page |
| Claude API ID (Mythos 5.1) | claude-mythos-5-1 | Model pages; Glasswing page |
| Claude API ID (Mythos 5.2) | claude-mythos-5-2 | Model pages; Glasswing page |
| Release date (Mythos Preview) | Announced 2026-04-07 (Project Glasswing) | Model pages; Glasswing page |
| Release date (Mythos 5) | 2026-06-09 | Model pages; Glasswing page |
| Release date (Mythos 5.1) | 2026-09-01 | Model pages; Glasswing page |
| Release date (Mythos 5.2) | 2026-11-09 | Model pages; Glasswing page |
| Status (Mythos Preview) | Deprecated 2026-06-09 | Model pages; Glasswing page |
| Status (Mythos 5, Mythos 5.1, Mythos 5.2) | Active (verification required) | Model pages; Glasswing page |
| Context window (Mythos Preview) | 1M (pricing page, standard rate) | Model pages; Glasswing page |
| Max output, modalities, thinking, default effort and knowledge cutoff (Mythos Preview) | Not stated by Anthropic on the pages checked (2026-11-15) | Model pages; Glasswing page |
| Context window (Mythos 5, Mythos 5.1, Mythos 5.2) | 1M tokens | Model pages; Glasswing page |
| Max output (Mythos 5, Mythos 5.1, Mythos 5.2) | 128K tokens | Model pages; Glasswing page |
| Modalities (in / out) (Mythos 5, Mythos 5.1, Mythos 5.2) | Text, image / text | Model pages; Glasswing page |
| Thinking (Mythos 5, Mythos 5.1, Mythos 5.2) | Adaptive, always on | Model pages; Glasswing page |
| Default effort (Mythos 5, Mythos 5.1, Mythos 5.2) | high | Model pages; Glasswing page |
| Knowledge cutoff (reliable / training) (Mythos 5) | Jan 2026 / Jan 2026 | Model pages; Glasswing page |
| Knowledge cutoff (reliable / training) (Mythos 5.1, Mythos 5.2) | Jun 2026 / Jun 2026 | Model pages; Glasswing page |
| Paired Fable model (Mythos Preview) | None | Model pages; Glasswing page |
| Paired Fable model (Mythos 5) | Claude Fable 5 | Model pages; Glasswing page |
| Paired Fable model (Mythos 5.1) | Claude Fable 5.1 | Model pages; Glasswing page |
| Paired Fable model (Mythos 5.2) | Claude Fable 5.2 | Model pages; Glasswing page |
| Open weights / licence | No / proprietary | Model pages; Glasswing page |
| Parameters | Not stated by Anthropic | Model pages; Glasswing page |
Release and version history
As of 2026-11-15. Source: Project Glasswing page and the Claude Platform release notes.
| Date | Event | API IDs |
|---|---|---|
| 2026-04-07 | Project Glasswing announced with Claude Mythos Preview for launch partners and over 40 further organisations that maintain critical software | claude-mythos-preview |
| 2026-06-09 | Claude Mythos 5 launched for Project Glasswing participants alongside Claude Fable 5; Mythos Preview deprecated | claude-mythos-5 |
| 2026-07-01 | Access to Claude Fable 5 and Claude Mythos 5 restored, with a link to Anthropic's statement on redeploying Fable 5 | claude-mythos-5 |
| 2026-09-01 | Claude Mythos 5.1 launched for Project Glasswing participants alongside Fable 5.1; cache reads cut to $0.25 per million tokens; text watermarking | claude-mythos-5-1 |
| 2026-10-06 | Cyber Verification Program expanded to three access tiers; existing Glasswing members move to the Specialized Access tier (Glasswing page update) | All Mythos models |
| 2026-11-09 | Claude Mythos 5.2 launched for organisations verified through Anthropic's verification programmes, with the same capabilities as Fable 5.2 | claude-mythos-5-2 |
Launch wording versus current wording: the release notes launched Mythos 5 and 5.1 "for Project Glasswing participants". The current model pages for both now state access through Anthropic's verification programmes, consistent with the October 2026 change that folded Glasswing members into a Cyber Verification Program tier.
Release-specific API behaviour: on Mythos 5 and later, disabled thinking, manual thinking budgets and assistant prefill are not supported. On Mythos 5.1, tool_choice types "any" and "tool" return a 400 error. Mythos 5.2 enforces the prefix check for every account, and no other model reads thinking blocks produced by Fable 5.2 or Mythos 5.2.
How much does Claude Mythos cost?
As of 2026-11-15. USD per million tokens (MTok), standard global rates. Source: Anthropic pricing page.
| Item | Price | Notes | Source |
|---|---|---|---|
| Mythos Preview | Input: $25 (after usage credits, Glasswing page); Output: $125 (after usage credits, Glasswing page); 5-minute cache write: Not stated by Anthropic; 1-hour cache write: Not stated by Anthropic; Cache read: Not stated by Anthropic; Batch (in / out): Not stated by Anthropic; Fast mode (in / out): Not stated by Anthropic | None | Pricing page; Glasswing page |
| Mythos 5 | Input: $10; Output: $50; 5-minute cache write: $12.50; 1-hour cache write: $20; Cache read: $1; Batch (in / out): $5 / $25; Fast mode (in / out): Not stated by Anthropic | None | Pricing page; Glasswing page |
| Mythos 5.1 | Input: $10; Output: $50; 5-minute cache write: $12.50; 1-hour cache write: $20; Cache read: $0.25; Batch (in / out): $5 / $25; Fast mode (in / out): Not stated by Anthropic | None | Pricing page; Glasswing page |
| Mythos 5.2 | Input: $8; Output: $40; 5-minute cache write: $10; 1-hour cache write: $16; Cache read: $0.20; Batch (in / out): $4 / $20; Fast mode (in / out): $16 / $80 (Claude API) | None | Pricing page; Glasswing page |
- Long context: the pricing page states that Claude 4.6 and later models, except Haiku 5.5, and Claude Mythos Preview include the full 1M-token window at standard pricing.
- US-only inference: setting inference_geo to US-only adds a 1.1x multiplier to every token category on Claude 4.6 and later models.
- Tokeniser: Claude 4.7 and later models and Mythos Preview use a newer tokeniser that produces roughly 30% more tokens for the same text than earlier models.
Context window and output limits
| Limit | Value | Notes | Source |
|---|---|---|---|
| Mythos Preview | Context window: 1M (pricing page); Max output: Not stated by Anthropic; Long-context surcharge: None (standard pricing) | None | Model pages; pricing page |
| Mythos 5 | Context window: 1M; Max output: 128K; Long-context surcharge: None (standard pricing) | None | Model pages; pricing page |
| Mythos 5.1 | Context window: 1M; Max output: 128K; Long-context surcharge: None (standard pricing) | None | Model pages; pricing page |
| Mythos 5.2 | Context window: 1M; Max output: 128K; Long-context surcharge: None (standard pricing) | None | Model pages; pricing page |
Modalities and tools
- Effort: per-message effort changes are in beta on Mythos 5.1 on the Claude API and on Google Cloud (3 September 2026 note).
- Code execution: code_execution_20260120 is available on Mythos 5 (18 June 2026 note).
- Computer use and browser use toolsets: available for Mythos 5 on the Claude API (19 August 2026) and Google Cloud (20 August 2026). For the 5.2 pair, requests declaring computer_20251124 return a 400 error on every platform except Amazon Bedrock.
- Advisor tool (beta): a 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor.
- Watermarking: text generated by Mythos 5.1 carries Anthropic's text watermark (1 September 2026 note).
Where is Claude Mythos available?
As of 2026-11-15. Availability on every platform is subject to the access conditions in the next section.
| Channel | Details | Source |
|---|---|---|
| Claude API (Mythos 5) | claude-mythos-5 | Model pages |
| Claude API (Mythos 5.1) | claude-mythos-5-1 | Model pages |
| Claude API (Mythos 5.2) | claude-mythos-5-2 | Model pages |
| Amazon Bedrock (Mythos 5) | anthropic.claude-mythos-5 | Model pages |
| Amazon Bedrock (Mythos 5.1) | anthropic.claude-mythos-5-1 | Model pages |
| Amazon Bedrock (Mythos 5.2) | global.anthropic.claude-mythos-5-2 | Model pages |
| Claude Models on Google Cloud (Google managed) (Mythos 5) | claude-mythos-5 | Model pages |
| Claude Models on Google Cloud (Google managed) (Mythos 5.1) | claude-mythos-5-1 | Model pages |
| Claude Models on Google Cloud (Google managed) (Mythos 5.2) | claude-mythos-5-2 | Model pages |
| Claude Models on Google Cloud (Anthropic managed) (Mythos 5, Mythos 5.1) | Not listed on the model page | Model pages |
| Claude Models on Google Cloud (Anthropic managed) (Mythos 5.2) | claude-mythos-5-2 | Model pages |
| Microsoft Foundry (Mythos 5) | claude-mythos-5 | Model pages |
| Microsoft Foundry (Mythos 5.1) | claude-mythos-5-1 | Model pages |
| Microsoft Foundry (Mythos 5.2) | claude-mythos-5-2 | Model pages |
For Mythos Preview, the Glasswing page states that participants can use the model through the Claude API, Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry. The Fable 5.2 and Mythos 5.2 system card adds that Mythos 5.2 powers Claude Security for all Claude Enterprise customers. For enterprise controls in the Claude apps see the Claude for Enterprise Security Guide.
Access restrictions
The Glasswing page states: "We do not plan to make Claude Mythos Preview generally available." Anthropic gives as its reason the need to build safeguards that detect and block the model's most dangerous outputs before deploying Mythos-class models at scale.
| Model | Who can use it (as stated by Anthropic) | How to request |
|---|---|---|
| Mythos Preview | Project Glasswing launch partners and over 40 further organisations that build or maintain critical software; open-source maintainers via the Claude for Open Source program | Glasswing participation; Claude for Open Source application |
| Mythos 5 | Launched for Glasswing participants; model page now states organisations verified through Anthropic's verification programmes | Apply to the verification programme that covers the use case |
| Mythos 5.1 | Launched for Glasswing participants; model page now states organisations verified through Anthropic's verification programmes | Apply to the verification programme that covers the use case |
| Mythos 5.2 | Organisations verified through Anthropic's verification programmes, such as the Cyber Verification Program; the system card also names the Life Sciences Verification Program | Apply to the verification programme that covers the use case |
The Glasswing launch partners named by Anthropic, alphabetically, are Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. Organisations that cannot obtain verification use the Fable model of the same version and can fall back to earlier models where the developer opts in.
Benchmarks (provider-reported, dated)
All figures are reported by Anthropic; no independent leaderboard results are included. Scores come from different evaluation set-ups and are not directly comparable across rows.
| Benchmark | Model | Score | Comparator in the same source | Source and date |
|---|---|---|---|---|
| CyberGym | Mythos Preview | 83.1% | Opus 4.6: 66.6% | Glasswing page, April 2026 |
| SWE-bench Verified | Mythos Preview | 93.9% | Opus 4.6: 80.8% | Glasswing page, April 2026 |
| SWE-bench Pro | Mythos Preview | 77.8% | Opus 4.6: 53.4% | Glasswing page, April 2026 |
| Terminal-Bench 2.0 | Mythos Preview | 82.0% | Opus 4.6: 65.4% | Glasswing page, April 2026 |
| GPQA Diamond | Mythos Preview | 94.6% | Opus 4.6: 91.3% | Glasswing page, April 2026 |
| Humanity's Last Exam (no tools / with tools) | Mythos Preview | 56.8% / 64.7% | Opus 4.6: 40.0% / 53.1% | Glasswing page, April 2026 |
| BrowseComp | Mythos Preview | 86.9% | Opus 4.6: 83.7% | Glasswing page, April 2026 |
| OSWorld-Verified | Mythos Preview | 79.6% | Opus 4.6: 72.7% | Glasswing page, April 2026 |
| Anthropic ECI (AECI) | Mythos 5.2 | 177.12 | Mythos 5.1: 172.62; Opus 5.5: 174.78 | Fable 5.2 and Mythos 5.2 system card, 9 Nov 2026 |
| ExploitBench, full arbitrary code execution | Mythos 5.2 | 90.0% (369 of 410) | Not stated in the same row | Fable 5.2 and Mythos 5.2 system card, 9 Nov 2026 |
| SWE-bench Pro | Fable 5.2 (same model, general configuration) | 93.0% | Fable 5.1: 81.2%; Opus 5.5: 89.9% | Fable 5.2 and Mythos 5.2 system card, 9 Nov 2026 |
Anthropic's Glasswing page notes that Mythos Preview may perform well on Humanity's Last Exam at low effort, which could indicate some memorisation.
Model card and system card
| Model | System card | Main points stated (summary) |
|---|---|---|
| Mythos Preview | Not linked from the pages checked | Glasswing page sets out capabilities, partners and the decision not to release generally |
| Mythos 5 | Claude Fable 5 and Mythos 5 system card | Safety evaluations and deployment decisions for Fable 5 and Mythos 5 (per the model page) |
| Mythos 5.1 | Claude Fable 5.1 and Mythos 5.1 system card | Safety evaluations and deployment decisions for Fable 5.1 and Mythos 5.1 (per the model page) |
| Mythos 5.2 | Claude Fable 5.2 and Mythos 5.2 system card, published 9 November 2026 | See below |
The Mythos 5.2 system card, as summarised from Anthropic's text:
- Chemical and biological risk: a modest improvement over prior models that does not cross the CB-2 threshold; CB-1 protections are applied.
- AI research and development: Anthropic judges the model does not double the rate of AI R&D progress, and calls this assessment uncertain.
- Alignment: Anthropic rates the risk of catastrophic harm from misalignment as low, while noting higher evaluation awareness, with white-box evidence of awareness in up to 43% of automated audit transcripts.
- Cyber filter on the general configuration: Fable 5.2 uses a three-stage filter (activation probe, lightweight classifier, LLM classifier); vulnerability discovery in source code is allowed and in compiled binaries is blocked. Mythos 5.2 is the configuration without that block for verified users.
For how these documents fit wider frontier safety practice see the Frontier AI Model Risk Frameworks Guide.
Data handling
- 30-day retention: the release notes state that Mythos 5.1 and Mythos 5.2 require 30-day data retention and are not available under zero data retention (ZDR) unless expressly authorised by Anthropic; Fable 5, the pair of Mythos 5, carries the same 30-day rule.
- Thinking blocks: on the 5.2 pair, thinking blocks work only in the account that produced them or an account linked to it.
For controls on the API keys and cloud roles that can reach Mythos, see the LLMjacking Guide.
Deprecation status
| Model or item | Details | Source |
|---|---|---|
| claude-mythos-preview | Status: Deprecated; Deprecated: 2026-06-09; Retirement: To be announced | Model deprecations |
| claude-mythos-5 | Status: Active; Deprecated: N/A; Retirement: Not sooner than 2027-06-09 | Model deprecations |
| claude-mythos-5-1 | Status: Active; Deprecated: N/A; Retirement: Not sooner than 2027-09-01 | Model deprecations |
| claude-mythos-5-2 | Status: Active; Deprecated: N/A; Retirement: Not sooner than 2027-11-09 | Model deprecations |
Anthropic states that "not sooner than" dates are the earliest possible retirement dates, not scheduled retirements. It gives at least 60 days' notice before retiring publicly released models and notifies customers with active deployments by email; the deprecations page does not say how that notice period applies to restricted models. Anthropic publishes migration guides from Mythos 5 and Mythos 5.1 to Mythos 5.2.
How Claude Mythos differs from other models in its family
As of 2026-11-15. Source: Mythos 5.2 model page comparison table.
| Model | Access | Input / output price | Default effort | Comparative latency |
|---|---|---|---|---|
| Claude Fable 5.2 | Generally available | $8 / $40 | high | Moderate |
| Claude Haiku 5.5 | Generally available | From $0.10 / $0.50 | medium | Fastest |
| Claude Mythos 5.2 | Verification required | $8 / $40 | high | Moderate |
| Claude Opus 5.5 | Generally available | $4 / $20 | medium | Moderate |
| Claude Sonnet 5.5 | Generally available | $2 / $10 | high | Fast |
- Versus Opus, Sonnet and Haiku: these are generally available lines at lower prices; see the Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 profiles.
- Across providers: Google also publishes an allowlisted cyber variant of its Gemini 4 Carbon model; see the Gemini 4 Carbon profile and Frontier AI Models Compared.
Practitioner checklist
- Confirm the organisation's verification status and programme tier before planning any workload on a Mythos model.
- Pin code to an explicit API ID (claude-mythos-5-2 or a cloud equivalent) and record it in the AI inventory.
- Remove any remaining calls to claude-mythos-preview.
- Check that the 30-day retention requirement fits your data commitments.
- Restrict which API keys and cloud roles can reach Mythos model IDs and review the Console usage export by key.
- Re-test prompts and tool calls on migration: tool_choice limits on 5.1 and computer use limits on the 5.2 pair differ by platform.
- Budget with the published rates, including the 1.1x US-only inference multiplier where used.
Standards and references
- Anthropic: Claude Mythos 5.2 model page (accessed 2026-11-15)
- Anthropic: Claude Mythos 5.1 model page (accessed 2026-11-15)
- Anthropic: Claude Mythos 5 model page (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Model deprecations (accessed 2026-11-15)
- Anthropic: Claude Platform release notes (accessed 2026-11-15)
- Anthropic: Project Glasswing (accessed 2026-11-15)
- Anthropic: Claude Fable 5.2 and Claude Mythos 5.2 system card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Anthropic Models, Claude Fable 5.2, Frontier AI Models Compared, Frontier AI Model Risk Frameworks Guide.