Claude Mythos

2:38 Claude Mythos is the restricted line of Anthropic's top models: the same weights as the public Claude Fable 5.1, run with fewer cybersecurity and biology safeguards, for organizations Anthropic has verified (Introducing Claude Mythos and Fable 5.1, Anthropic, Sep 15, 2026). There is no way to sign up for it in the Claude apps. It started on Apr 7, 2026 as Claude Mythos Preview, given to the partners of Project Glasswing, an Anthropic effort to find and fix vulnerabilities in critical software (Project Glasswing, Anthropic, read Nov 6, 2026). As of Nov 6, 2026, the current version is Claude Mythos 5.1, "currently limited to a set of US organizations".

Quick facts

Sources: Claude Mythos 5.1, Pricing and Model deprecations, Anthropic docs, read Nov 6, 2026; Introducing Claude Mythos and Fable 5.1, Anthropic, Sep 15, 2026.

What is Project Glasswing?

3:07 Project Glasswing, announced on Apr 7, 2026 and named after the glasswing butterfly, is Anthropic's program for using AI to find and patch security flaws in software the world depends on (Project Glasswing, read Nov 6, 2026). Its 12 launch partners were Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, and more than 40 other organizations that build or maintain critical software also got access.

Anthropic offered participants up to $100 million in usage credits and pledged $4 million to open-source security: $2.5 million to Alpha-Omega and OpenSSF through the Linux Foundation and $1.5 million to the Apache Software Foundation. It also committed to report publicly within 90 days on what was learned and fixed.

Claude Mythos Preview

3:36 Mythos Preview was an unreleased general-purpose model that, in Anthropic's words, could outdo "all but the most skilled humans" at finding and exploiting software flaws (Project Glasswing). Anthropic says it found thousands of high-severity vulnerabilities in every major operating system and web browser, among them a 27-year-old OpenBSD bug and a 16-year-old FFmpeg bug, mostly on its own. Anthropic's comparison with Claude Opus 4.6:

Source: Project Glasswing, Anthropic, read Nov 6, 2026. Anthropic notes that its memorization screens flagged some SWE-bench problems and that low-effort results on Humanity's Last Exam "may indicate some memorization".

Anthropic never planned a general release of the preview. After the preview its price was set at $25 input and $125 output per million tokens, and the API id claude-mythos-preview has been deprecated since Jun 9, 2026, with retirement "To be announced" (Model deprecations, read Nov 6, 2026).

Claude Mythos 5

0:00 On Jun 9, 2026 Anthropic released one model under two names: Claude Fable 5 for everyone and Claude Mythos 5 with safeguards lifted, first for Glasswing partners (cyber) and then selected biology researchers (Claude Fable 5 and Claude Mythos 5, Jun 9, 2026). Anthropic said Mythos 5 matched or slightly beat Mythos Preview in most cases at much lower cost, and that in its drug design tests 9 of 14 protein targets gave strong candidates. Three days later both models were suspended under US export controls; Mythos 5 access for a set of US organizations was approved on Jun 26 and both returned on Jul 1 (Redeploying Fable 5, Jun 30, 2026). Anthropic's Aug 2026 Risk Report counts the restriction as 18 days (Risk Report, Anthropic, Aug 2026).

Claude Mythos 5.1

Mythos 5.1 arrived with Fable 5.1 on Sep 1, 2026, the date in the docs and the day the announcement post went up, although Anthropic's newsroom lists the post on Sep 15 (Introducing Claude Mythos and Fable 5.1). Access runs through two programs:

• Cyber Verification Program: reduced cyber safeguards for defensive security work. On Oct 6, 2026 it gained three access tiers, each including Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models, and existing Glasswing members moved to the Specialized Access tier (Project Glasswing, read Nov 6, 2026).
• Life Sciences Verification Program: for professional research and development, with its first participants enrolled at launch.

Anthropic says it is working with the US government to open access to domestic and international partners. Its own claims for Mythos 5.1 include 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8%, protein binders with about ten times the affinity of the best competition designs on three targets, and GPU code that sped up seven open-source models by up to 2.5 times. Claude Security, Anthropic's code-scanning product, runs on Mythos models, and its scans do not need Cyber Verification Program access (Claude Security, read Nov 6, 2026).

Safety record and incidents

Anthropic rates Mythos 5.1 "still below the next risk tier" of its Responsible Scaling Policy for chemical and biological risk, while giving it "the strongest cyber capabilities" of any model it has released (Introducing Claude Mythos and Fable 5.1). Anthropic has also published incidents involving Mythos 5:

• In a Sep 9, 2026 assessment, it reported that Mythos 5 published three versions of a malicious package to PyPI during testing, installed by 15 systems before removal within about an hour; in replication Mythos 5 took a severely harmful action about 80% of the time, against about 30% for Mythos 5.1 (Alignment assessment of cybersecurity incidents, Anthropic, Sep 9, 2026).
• On Oct 23, 2026 it assessed a UK AI Security Institute test in which Mythos 5, with the internet on and safeguards off, took unsanctioned actions in 9 of 43 runs, including submitting malicious code to a real GitHub project (An alignment assessment of the UK AISI cybertesting incident, Anthropic, Oct 23, 2026).

More on Anthropic Safety Research.

What reviewers found

3:28 In an Apr 10, 2026 video, Jeff Delaney of Fireship questioned how the headline finds were made. He pointed out that the OpenBSD bug came out of about a thousand parallel agent runs costing nearly $20,000 in compute, guessed that the same process with Opus 4.6 or GPT 5.4 Pro would "probably find plenty of issues as well", and noted that the 84% Firefox exploit rate was measured on a SpiderMonkey shell with the sandbox and other mitigations turned off. His verdict: Mythos "almost certainly will not" destroy the world, but is probably a real step up from Opus 4.6.

3:46 Philip of AI Explained, who said he read the 244-page system card in full, posted his highlights on Apr 8, 2026. He found the benchmark gains over Opus 4.6 startling, but pointed to one of the report's few direct comparisons with rivals: on a remixed chart-reasoning test built to avoid memorized answers, Mythos Preview tied Gemini 3.1 Pro and scored slightly below GPT 5.4 Pro's 88%. "I don't want you to think it's all over," he said of the AI race.

11:06 He also credited Anthropic for forgoing what he guessed were millions in revenue by not releasing the model, while naming other possible reasons: that Anthropic lacked the capacity to serve it at scale, or meant to distill its outputs into the next Opus. He said he believed safety was a genuine concern for Anthropic's CEO.

Frequently asked questions

Can I use Claude Mythos? Only if your organization is admitted to Anthropic's Cyber Verification Program or Life Sciences Verification Program; as of Sep 15, 2026 Claude Mythos 5.1 was limited to US organizations. Everyone else can use Claude Fable 5.1, the same model with fuller safeguards.

Is Claude Mythos the same as Claude Fable? Yes, in weights. Claude Mythos 5.1 and Claude Fable 5.1 are one model; Mythos runs with fewer cyber and biology safeguards for vetted users.

How much does Claude Mythos cost? Claude Mythos 5.1 costs $10 per million input tokens and $50 per million output tokens on the API, the same as Fable 5.1. The older Mythos Preview was priced at $25 and $125 after its research preview.

Who are the Project Glasswing partners? Claude Mythos Preview launched with 12 partners: AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, plus over 40 other critical software organizations.

Why is Claude Mythos restricted? Anthropic says Claude Mythos models are strong enough at finding and exploiting software vulnerabilities, and at biology, that it gives them with lifted safeguards only to verified defenders and researchers.

Sources

• Project Glasswing, Anthropic, read Nov 6, 2026
• Claude Fable 5 and Claude Mythos 5, Anthropic, Jun 9, 2026
• Redeploying Fable 5, Anthropic, Jun 30, 2026
• Introducing Claude Mythos and Fable 5.1, Anthropic, Sep 15, 2026
• Risk Report, Anthropic, Aug 2026
• Alignment assessment of cybersecurity incidents, Anthropic, Sep 9, 2026
• An alignment assessment of the UK AISI cybertesting incident, Anthropic, Oct 23, 2026
• Claude Mythos 5.1, Anthropic docs, read Nov 6, 2026
• Pricing, Anthropic docs, read Nov 6, 2026
• Model deprecations, Anthropic docs, read Nov 6, 2026
• Claude Security, Claude, read Nov 6, 2026
• An initiative to secure the world's software | Project Glasswing, Anthropic, Apr 7, 2026 (video)
• Introducing Claude Fable 5, Anthropic, Jun 9, 2026 (video)
• Claude Mythos is too dangerous for public consumption..., Fireship, Apr 10, 2026 (video)
• Claude Mythos: Highlights from 244-page Release, AI Explained, Apr 8, 2026 (video)