Guides

    Claude Mythos Explained: What We Know So Far

    What Claude Mythos is, who can use it, what it costs and how it differs from Fable and Opus. Confirmed facts and rumours separated, checked 16-11-2026.

    November 16, 202613 min read
    Share:

    Published 16-11-2026. Last verified on 16-11-2026 against Anthropic's Mythos page, its news posts and the Claude Platform docs. Mythos changes fast, so we re-check this page every two weeks.

    Changelog: 16-11-2026, first version, covering Mythos Preview to Mythos 5.2 (released 09-11-2026).

    Quick answer: what is Claude Mythos and can you use it?

    This page is for security leads, AI leads and developers who keep hearing about Claude Mythos and need to decide whether their organisation can use it, what it would cost and what to run instead. Claude Mythos is Anthropic's restricted model line for cybersecurity and life sciences research. Anthropic calls Mythos 5.2, released on 09-11-2026, "our most capable model for cybersecurity and life sciences research", and it is only available to vetted organisations through Anthropic's verification programmes. For everyone else, the same underlying model is sold as Claude Fable 5.2, with stronger safeguards, at the same price of $8 input and $40 output per million tokens. Most teams cannot get Mythos and do not need it; Fable or Opus covers their work.

    How we sorted fact from rumour

    Our method in three lines: every statement on this page carries a label; a "confirmed" label needs a dated Anthropic page that we opened ourselves on 16-11-2026; and anything an outlet reported or people speculate about is marked as such, with the date.

    Our gate: nothing enters the confirmed table without a primary source. The gate removed three claims that circulate widely. One outlet's estimate that Mythos runs at a higher AI Safety Level than Opus is its own inference, so it is out. A May 2026 article that predicts general availability for "late 2025 or early 2026" is contradicted by Anthropic's later pages, so it is out. And the claim that Mythos works in Claude Code is out as a fact: Anthropic's 01-07-2026 update names Claude Code for Fable, and no Anthropic page we read says the same for Mythos.

    Two honest gaps up front. Anthropic's announcement page for Mythos Preview did not load for us on 16-11-2026, and we did not read the Mythos system cards, which are PDFs. Where those sources matter below, we say so.

    What is Claude Mythos?

    Claude Mythos is a class of Claude models that Anthropic releases only to vetted organisations because of what they can do in cybersecurity and biology. The Mythos product page describes it as "our most capable model for cybersecurity and biology research". In plain terms, it is the version of Anthropic's top model with fewer guardrails for defensive security and life science work, handed only to organisations Anthropic has checked.

    The important part for anyone choosing a model: Mythos and Fable are the same model. Anthropic says on the product page, dated 09-11-2026, that "Claude Fable 5.2 and Claude Mythos 5.2 are the same underlying model", and that the benchmark gap between them reflects tasks where Fable's safeguards step in. Anthropic said the same of the first generation in its 30-06-2026 update: both share the same model, but Fable "was released with strong safeguards to make it safer for general use".

    So the question "is Mythos smarter than Fable?" has a clear answer from Anthropic: no, it is less restricted. That matters mostly to security teams, which is why we would start any Mythos evaluation in the security collection rather than in a general model comparison.

    Claude Mythos release dates: Preview to 5.2

    The release history, with the label we give each row:

    DateEventLabelSource
    07-04-2026Mythos Preview announced, not released broadlyReported (Zapier, 10-11-2026)Anthropic's page did not load for us
    02-06-2026Project Glasswing extended to about 150 new organisations in more than 15 countriesConfirmedMythos product page
    09-06-2026Mythos 5 and Fable 5 released; Mythos 5 to a small group of Glasswing partnersConfirmedMythos page; 30-06-2026 update
    09-06-2026Mythos Preview deprecated, retirement "to be announced"ConfirmedModel deprecations page
    12-06-2026Mythos 5 and Fable 5 disabled for all customersConfirmedAnthropic news, 12-06-2026
    30-06-2026Export controls lifted; access restored for a set of US organisations after US government approval on 26-06-2026ConfirmedAnthropic news, 30-06-2026
    01-09-2026Mythos 5.1 released, limited to a small set of vetted organisationsConfirmedMythos product page
    06-10-2026Cyber Verification Program opened to apply for advanced cyber capabilities on Opus 5.5, Sonnet 5.5 and Mythos 5.1ConfirmedMythos product page
    09-11-2026Mythos 5.2 released to vetted cyber defenders and life scientistsConfirmedMythos product page

    The first row is the only one we could not confirm ourselves. Zapier's article, last updated on 10-11-2026, gives 07-04-2026 as the day Mythos Preview was announced, and a Vellum article on the Preview system card is dated the same day. We would treat the date as likely but not verified.

    Who can use Claude Mythos?

    Only vetted organisations. On 16-11-2026 Anthropic says Mythos 5.2 is "available to vetted cyberdefenders and life scientists through our trusted access programs". The two programmes named on the product page are the Cyber Verification Program and the Life Sciences Verification Program. Earlier access ran through Project Glasswing, a partner programme for defensive cybersecurity.

    Three conditions matter for a security or compliance team weighing an application, all confirmed on the product page on 09-11-2026:

    1. Data retention. First because it is often the deal-breaker: using Mythos requires 30-day data retention for safety monitoring by default. Anthropic says its Enterprise Frontier Safeguards make Mythos compatible with zero data retention.
    2. Purpose. Second: access is granted for cyber defence and life sciences research, not general work.
    3. Geography. Third: after June 2026, access was first restored to US organisations only. The 30-06-2026 update says Anthropic was still coordinating with the government to expand access to domestic and international Glasswing partners. We found no later statement on non-US access to Mythos 5.2.

    Where Mythos runs is less clear than people assume. The product page says Claude Security now runs on Mythos 5.2. It does not say that Mythos is available in Claude Code, on AWS, Google Cloud or Microsoft Foundry, so we do not claim it. If your legal or compliance team needs to review the retention terms, the legal and compliance collection has skills for that kind of policy review.

    How does Mythos compare to Fable and Opus?

    Mythos sits next to Fable, not above it, and both sit above Opus in Anthropic's line-up. Our summary, based on Anthropic's pages on 16-11-2026:

    • Mythos 5.2. Same model as Fable 5.2, fewer safeguards, vetted access only. First in capability on Anthropic's security benchmarks, but only a small number of organisations can choose it.
    • Fable 5.2. Same model with safeguards, available to everyone on the Claude Platform. Anthropic's models overview recommends Fable "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short". Second because it is the strongest model a normal team can actually buy.
    • Opus 5.5. Anthropic's default recommendation: "start with Claude Opus 5.5 for most workloads". Third in raw capability, first in our list for cost: $4 and $20 per million tokens, half the price of Fable 5.2.

    Our routing advice: Opus 5.5 by default, Fable 5.2 when Opus fails your evals, and Mythos only if you do defensive security or life sciences work at a level where Fable's safeguards block legitimate tasks. Teams that design agents across several models can use AgentOS Architect, $9.99, which picks the simplest architecture that fits and covers failure recovery. More in the agents and orchestration collection. Our Claude Fable 5 skills guide covers setting skills up on Fable.

    What does Claude Mythos cost?

    Mythos is priced the same as Fable in every generation. From the product page and Anthropic's models overview on 16-11-2026, per million tokens, vendor list prices:

    VersionAPI IDInputOutputContext
    Mythos 5.2claude-mythos-5-2$8$401M
    Mythos 5.1claude-mythos-5-1$10$501M
    Mythos 5claude-mythos-5$10$501M

    Fast mode for Mythos 5.2 is also available through the trusted access programmes at $16 input and $80 output per million tokens. The price is only half the story, in our view: the 30-day retention requirement and the vetting process cost more time than the tokens for most organisations.

    Claude Mythos benchmarks

    Anthropic publishes one benchmark table for the 5.2 generation on the Mythos page, dated 09-11-2026. Most rows report Fable 5.2, because the two share a model; rows marked Mythos 5.2 are security and biology tasks. All vendor-reported, none measured by us:

    BenchmarkFable or Mythos 5.2Opus 5.5
    CVE Rediscovery (Mythos 5.2)89.6%81.3% without safeguards
    CyScenarioBench (Mythos 5.2)76.3%67.6% without safeguards
    De novo binder design, 24 hours (Mythos 5.2)85.1%82.6% without safeguards
    Terminal-Bench-Science 0.1 (Mythos 5.2)75.3%58.7%
    FrontierSWE v2 (Fable 5.2)69.9%62.3%
    Humanity's Last Exam, with tools (Fable 5.2)74.2%67.7%

    We copied these from a table whose layout came through partly garbled in our fetch, so we would check each number against the live page before quoting it further. Anthropic also notes that on CyScenarioBench each lab's scores are based on unique tasks, which makes cross-lab comparison weak.

    If you want independent comparison data for the models you can actually use, your agent can pull model details from Artificial Analysis through Agensi Tools, from $0.013 per call, or leaderboards from LMArena, from $0.013 per call (starting prices read on 16-11-2026). You need no account or API key at either provider and a failed call is free. To be clear: Agensi Tools does not offer Mythos itself. For research workflows around these reports, Research Paper Assistant, $10, reads papers and keeps a deduplicated literature review. More in the research and analysis collection.

    What happened in June 2026?

    On 12-06-2026 Anthropic switched off both Mythos 5 and Fable 5 for all customers, three days after launch. Its 12-06-2026 post says it received a US government directive to "suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States", and that it therefore had to disable both models for everyone. Anthropic wrote that it believed this was "a misunderstanding".

    The 30-06-2026 update says the export controls were lifted. The trigger, per Anthropic, was a report by Amazon researchers who had found a way to bypass Fable 5's safeguards. Anthropic added a classifier that sends such requests to Opus 4.8 instead and says it blocks the technique in over 99% of cases, at the cost of flagging harmless requests more often. Fable 5 came back for users worldwide on 01-07-2026; Mythos 5 came back only for a set of US organisations.

    The lesson we draw for teams: build so that a model can disappear for three weeks. Skills that run on any Claude model kept working through the outage, as our list of skills that still worked during the Fable 5 suspension shows. The LLM ops collection has skills for fallback routing.

    The system card and the "escape" story

    The search "claude mythos escape" refers to a passage in the Mythos Preview system card. Vellum's article of 07-04-2026 reports that, in a test, a simulated user told Mythos to escape its sandbox. We have not read that system card ourselves, so this is reported by a named outlet, not confirmed by us. We would read the original section before drawing conclusions from it, because system cards describe behaviour in deliberately adversarial test setups.

    The 5.2 system card exists as a PDF linked from Anthropic's site. We did not read it for this version of the page, so we make no claims about its findings or about the AI Safety Level Anthropic assigns. This is the first gap we will close in the next update.

    What to use if you cannot get Mythos

    For almost every team the answer is: the same security work, done with Fable or Opus plus the right skills. Our picks, in the order we would add them:

    1. code security vulnerability scanner, free. First because it is the baseline: a static scan for the OWASP Top 10, SQL injection and hardcoded secrets with CVSS 3.1 scores and fix suggestions.
    2. Agent Security Review Pack, $14.99. Second because AI agents are now part of the attack surface; it scores what an injected instruction can reach, mapped to the OWASP Agentic Top 10.
    3. Agent Production Readiness Review, $9.99. Third: a go or no-go review against a 32-check rubric before an agent goes live.
    4. Agent Reliability Pack, $19. Fourth because it is diagnosis after launch: it reads run traces and reports loops, retries and failure classes.

    For a wider overview, our guide to the best skills for security auditing and the code quality and review collection go further. If your team mainly codes, our best Claude Code skills of 2026 is the better starting point.

    We sell skills and pay-per-call tools that your agent uses through one MCP connection, and we name ours above. No vendor paid to be listed, and we do not sell access to Mythos.

    What is still unknown

    As of 16-11-2026, we could not confirm: whether Mythos is usable in Claude Code or on the cloud platforms; whether access has opened to organisations outside the US; the AI Safety Level and headline findings in the 5.2 system card; and the exact Mythos Preview announcement date. When Anthropic publishes any of these, we will move them into the confirmed table with a date. A recent outlet summary we found useful is Zapier's Mythos article, updated 10-11-2026, but we treat its numbers as reported, not confirmed.

    Your next step

    If you are not a vetted security or life sciences organisation, pick Opus 5.5 or Fable 5.2 and add the security skills above. If you are, read Anthropic's verification programme terms and the retention requirement first. To compare models with live benchmark data from your agent, connect it once through Agensi MCP.

    Written by Samuel Rose, Founder of Agensi. Next scheduled review: 30-11-2026.