AI/Explainer

What Is Claude Mythos 5.2? Anthropic's Cybersecurity Model Explained

Anthropic's gated, safeguard-free sibling to Claude Fable 5.2 finds vulnerabilities faster than any earlier Claude model — but you can't just sign up for it.

Anthropic logo on a dark background
Image: Anthropic.

Anthropic's Claude Mythos 5.2 is the company's most capable model for cybersecurity work: it finds, reproduces and fixes software vulnerabilities at a higher rate than any earlier Claude model, and it's built for "the most ambitious, open-ended technical work" rather than everyday chat. Released November 9, 2026 alongside Claude Fable 5.2, Mythos 5.2 is the same underlying model as Fable but without Fable's cyber and biology safeguards — which is why access to it is restricted to vetted security professionals and enterprise customers, not available to the general public the way a normal Claude model is.

Claude Mythos 5.2 at a glance

  • Released: November 9, 2026
  • Built for: cybersecurity engineering and other high-end technical/research work
  • Relationship to Fable 5.2: same base model, minus Fable's cyber and biology safeguards
  • Access: Cyber Verification Program, Life Sciences Verification Program, or Claude Enterprise via Claude Security
  • Price: $8 / $40 per million input/output tokens (standard); $16 / $80 in fast mode
  • Context window: 1M tokens; max output: 128K tokens
  • CVE Rediscovery benchmark: 89.6%, up from 78.3% for Mythos 5.1

What Claude Mythos 5.2 actually is

Mythos isn't a separate model architecture from Fable — Anthropic says the two are literally the same underlying model, trained together and released on the same day. The difference is what's layered on top. Claude Fable 5.2 ships with Anthropic's standard safeguards against misuse in cybersecurity and biology, making it suitable for general release. Claude Mythos 5.2 strips those specific guardrails out, which is what lets it dig into exploit development, vulnerability reproduction and other security research tasks that the guarded version is designed to decline. Anthropic's own framing is that Mythos models are meant for "the most ambitious, open-ended technical work" — the kind of unrestricted technical reasoning that a defanged, consumer-safe model can't fully do.

That tradeoff is also why you can't just sign up for Mythos 5.2 the way you'd subscribe to Claude Pro. Anthropic gates it behind verification: security researchers and companies apply through the Cyber Verification Program, life sciences researchers go through the separate Life Sciences Verification Program, and Claude Enterprise customers get indirect access through Claude Security, a product that's powered by Mythos 5.2 under the hood even for organizations that haven't individually applied for cyber verification. Anthropic's rationale is straightforward: a model this good at finding and exploiting vulnerabilities is also a model that could do real damage if it were available to anyone with a credit card, so it ties the capability to an identity-verified, legitimate-use context instead.

The cybersecurity benchmarks behind the "most capable" claim

Anthropic backs up the cybersecurity framing with internal benchmark numbers comparing Mythos 5.2 against its predecessor and against Claude Opus 5.5 running without any cyber safeguards at all:

BenchmarkMythos 5.1Opus 5.5 (no safeguards)Mythos 5.2
CVE Rediscovery78.3%81.3%89.6%
CyScenarioBench61.7%67.6%76.3%
Internal CVE Rediscovery (vs. other models)—+10.2% more bugs found by Mythos 5.2+14.4% more than Mythos 5.1

CVE Rediscovery tests whether a model, given a codebase with a known, previously disclosed vulnerability, can independently find that same bug without being told where to look — a proxy for real-world vulnerability hunting. CyScenarioBench evaluates more open-ended security scenarios. On both, Mythos 5.2's jump over Mythos 5.1 is bigger than the gap between Mythos 5.1 and an unguarded Opus 5.5, which is why Anthropic is pitching this specifically as a step up in practical, hands-on security capability rather than a general reasoning upgrade.

Two more claims from Anthropic's launch materials are worth flagging because they speak to reliability rather than raw capability. First, a security company found that one in six confirmed bugs it identified using Mythos 5.2 had been missed by every other frontier model it tested against — a sign the model is finding things competitors aren't, not just finding the same things faster. Second, one tester reported that earlier Claude models overstated roughly 1 in 20 vulnerability findings (false positives dressed up as real bugs), while Mythos 5.2 overstated none in that tester's evaluation — a meaningful trust signal for a tool meant to be used by professionals who have to act on its output.

What early users are saying

Anthropic's announcement includes on-record reactions from two named technical leads at companies using Mythos 5.2. Mozilla's Distinguished Engineer Brian Grinstead said the model "roughly doubled the vulnerability discovery metrics of Mythos 5.1 across respective scanning campaigns" — a specific, measurable claim rather than generic praise. Rubrik co-founder and CTO Arvind Nithrakashyap focused on usability over raw numbers, saying the output quality improvement meant "the output is more readable and empowered our team to spot-check faster and more easily." That second point matters in security work specifically: a model that finds more bugs is only useful if a human reviewer can quickly verify which findings are real before acting on them.

The life sciences side: why the same model has two verification programs

Cybersecurity isn't the only domain Anthropic pulled Fable 5.2's safeguards out for. The Life Sciences Verification Program exists because the same underlying capability jump that makes Mythos better at finding software vulnerabilities also makes it better at open-ended biological research tasks — and biology research, like security research, is an area where a model's raw capability and its potential for misuse rise together. Anthropic's launch benchmarks include a "de novo binder design" test, scored over a 24-hour task window, where Mythos 5.2 hit 85.1, up from 79.1 for Mythos 5.1 and 82.6 for an unguarded Opus 5.5. A separate benchmark, Terminal-Bench-Science 0.1, put Mythos 5.2 at 75.3%. Both numbers describe the model's ability to carry out long, open-ended scientific reasoning tasks without a human walking it through each step — the same trait that makes it useful for a security researcher chasing down an exploit chain translates directly into usefulness for a biologist designing a novel molecule, which is why Anthropic built one verification pipeline for cyber work and a structurally identical second one for life sciences rather than trying to write a single policy that covers both.

The common thread across both programs is identity and intent verification rather than a technical capability switch. Mythos 5.2 itself doesn't know which program granted a given user access — the model behaves the same way regardless of whether the request came through Cyber Verification or Life Sciences Verification. What changes is who Anthropic has agreed to let make requests at all, which is a policy and process control sitting outside the model rather than inside it.

Why gate a model instead of just making it safer

It's worth sitting with why Anthropic chose this specific structure — one base model, two safeguard profiles, with the unguarded profile locked behind identity verification — instead of simply releasing a single model that's "safe enough" for everyone. The honest answer is that cybersecurity capability and cybersecurity misuse potential aren't cleanly separable in the way, say, a chatbot's helpfulness and its tendency to hallucinate are. A model that's extremely good at finding a buffer overflow in a piece of unfamiliar code is, by construction, also extremely good at finding that same bug for someone who wants to exploit it rather than patch it. Anthropic's bet is that you get more net benefit — more real vulnerabilities found and fixed, faster — by giving the full, unguarded capability to vetted professionals than you would by giving everyone a weaker, hedged version that's safer in the median case but worse at the job in every case. Whether that bet pays off at scale is something only time and a track record of how the verification programs are actually used will answer, but it's a different model than the "one safety profile fits all users" approach most consumer AI products still take.

Pricing, context window and how to actually get access

Mythos 5.2 and Fable 5.2 share identical list pricing: $8 per million input tokens and $40 per million output tokens in standard mode, rising to $16 and $80 per million in Anthropic's faster-output mode — though fast mode for Mythos is itself restricted to organizations already inside one of the verification programs. Anthropic's model documentation lists a 1-million-token context window and a 128,000-token maximum output, a knowledge cutoff of June 2026, and adaptive thinking that's enabled by default at a high effort setting rather than something a developer has to turn on. Anthropic has committed to not retiring the model before November 9, 2027, giving security teams at least a year of stability to build workflows around it.

In practice, there are three doors into Mythos 5.2:

  • Cyber Verification Program: for individual security researchers and firms doing offensive or defensive security work who apply directly to Anthropic.
  • Life Sciences Verification Program: a parallel track for biology/life-sciences researchers, since the same base model without safeguards is also used for that kind of open-ended scientific work.
  • Claude Enterprise + Claude Security: the path most companies will actually use — Claude Security runs on Mythos 5.2 for Enterprise customers automatically, without those customers needing to separately pass cyber verification.

There's no self-serve signup for raw Mythos 5.2 access the way there is for Claude.ai or the standard API — this is a deliberately gated release, not a product tier you can buy your way into. If you're trying to decide which of the two models actually fits your use case, see our side-by-side comparison of Claude Fable 5.2 and Mythos 5.2.

How Anthropic frames the safety tradeoff

Removing cyber and biology safeguards from a frontier model is the kind of decision Anthropic has to justify on safety grounds, and its launch materials include alignment numbers alongside the capability ones. Mythos 5.2 scored 2.67 on an automated behavioral audit (lower is better), compared with 2.79 for Opus 5.5 and 3.10 for Sonnet 5.5. In a sandbox escape evaluation — testing whether a model will use planted credentials or files it shouldn't — Mythos 5.2 did so in 3.1% of scenarios, close to Opus 5.5's 3.3% and better than Sonnet 5.5's 5.1%. Anthropic's conclusion, in its own words, is that the risk of unsafe behavior from Mythos 5.2 in high-stakes situations is low. Those are Anthropic's own evaluations rather than independent third-party audits, which is a reasonable thing to keep in mind when weighing the claim, but the company is at least publishing comparative numbers rather than asserting safety without evidence.

What's next

For most Pandromeda readers, Mythos 5.2 won't be something you log into directly — it's infrastructure that shows up indirectly, through Claude Security inside Claude Enterprise, or through the work of verified security researchers who use it to find and report vulnerabilities faster. The more immediate thing to watch is whether other frontier labs follow Anthropic's structure here: a single base model split into a guarded public version and a gated, unguarded "pro" version for vetted specialists, rather than one model trying to serve both audiences with a single set of safeguards. If that CVE Rediscovery gap keeps widening with each release, expect verified-access programs like this one to become a standard feature of how AI companies handle security tooling, not an Anthropic-specific experiment.

Frequently asked questions

What is Claude Mythos 5.2?

Claude Mythos 5.2 is Anthropic's most capable AI model for cybersecurity work, released November 9, 2026. It's the same underlying model as Claude Fable 5.2 but without Fable's cyber and biology safeguards, which lets it handle vulnerability research and exploit reproduction that the guarded version is designed to decline.

How is Claude Mythos 5.2 different from Claude Fable 5.2?

They're the same base model, trained together and released on the same day. Fable 5.2 ships with Anthropic's standard cyber and biology safeguards for general use. Mythos 5.2 has those specific safeguards removed and is gated behind identity verification instead.

How much does Claude Mythos 5.2 cost?

Standard pricing is $8 per million input tokens and $40 per million output tokens — the same as Fable 5.2. A faster-output mode costs $16/$80 per million tokens, but fast mode is limited to organizations already verified through one of Anthropic's access programs.

Who can access Claude Mythos 5.2?

Access runs through three paths: the Cyber Verification Program for security researchers, the Life Sciences Verification Program for biology researchers, and Claude Enterprise customers, who get indirect access automatically through Claude Security without separately applying for cyber verification.

What is the CVE Rediscovery benchmark Anthropic cites?

It tests whether a model can independently find a known, previously disclosed software vulnerability in a codebase without being told where to look. Mythos 5.2 scored 89.6%, up from 78.3% for Mythos 5.1 and 81.3% for an unguarded Claude Opus 5.5.

When was Claude Mythos 5.2 released, and when will it retire?

It launched November 9, 2026 alongside Claude Fable 5.2. Anthropic has committed to not retiring the model before November 9, 2027.

Sources

More on Claude Mythos 5.2 →Claude Mythos 5.2AnthropicAI cybersecurityClaude Fable 5.2
Theo Park
Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from AI

See all