Claude Mythos Preview: April 2026 Debut, Strong Benchmarks, No Public Access

The short version

• Public debut: April 7, 2026, alongside Project Glasswing; Anthropic’s system-card index also lists Mythos Preview under April 2026.
• Performance: In Anthropic’s published comparison, Mythos Preview beats Claude Opus 4.6 on GPQA Diamond, Humanity’s Last Exam, BrowseComp and OSWorld-Verified.
• Access: It is restricted to partner organizations maintaining important software infrastructure, with terms limiting use to cybersecurity.

Release date: April 7, 2026 — but not a general launch

If you ask when Claude Mythos Preview was released, there are two different answers.

The public reveal was on April 7, 2026. Anthropic’s Project Glasswing page carries that date and says the initiative is powered by Claude Mythos Preview; Anthropic’s model system-card directory lists Mythos Preview in April 2026.

The general-availability answer is different. In software, general availability usually means a product is broadly open to ordinary customers or developers. Anthropic says it did not release Claude Mythos Preview for general availability.

What it is: a restricted model for critical software security

Project Glasswing is not a standard Claude product launch. Anthropic describes it as an initiative to secure the world’s most critical software for the AI era, powered by its most capable model yet, Claude Mythos Preview.

That security-first framing matters. Anthropic’s system card says the model’s capabilities can be used defensively, such as finding and fixing vulnerabilities in software code, but also offensively, such as designing sophisticated exploit methods. Anthropic says those capabilities are a major reason it chose not to make Mythos Preview generally available.

Anthropic’s red-team blog gives the same basic rationale: Mythos Preview was initially released through Project Glasswing to a limited group of critical industry partners and open-source developers so defenders could start securing important systems before models with similar capabilities become broadly available.

Benchmark results: ahead of Claude Opus 4.6 in Anthropic’s table

Anthropic’s published side-by-side comparison shows Claude Mythos Preview ahead of Claude Opus 4.6 on every listed benchmark:

Anthropic also says Mythos Preview scored higher on BrowseComp while using 4.9× fewer tokens than Opus 4.6.

Taken at face value, those are strong results. The safest reading is that Anthropic is presenting Mythos Preview as a frontier model with notable strength on demanding evaluations involving reasoning, tool use, browsing and computer-environment tasks.

Why the benchmark numbers need caveats

The scores are impressive, but they should not be stretched into a blanket claim that Mythos Preview is better in every real-world situation.

First, the direct benchmark comparison available here is Anthropic’s own published table rather than an independent reproduction. That does not make the numbers meaningless, but it does mean readers should treat them as official disclosure, not outside validation.

Second, Anthropic adds a caveat to the Humanity’s Last Exam result: Mythos still performs well on HLE at low effort, which could indicate some level of memorization. In other words, the HLE score is evidence for performance on that benchmark, not a complete guarantee of equal gains on unseen tasks.

Third, Anthropic’s alignment-risk update is more mixed on some internal evaluations. It says Mythos Preview was either slightly above or slightly below Claude Opus on Minimal-LinuxBench depending on the setting, and that SHADE-Arena stealth success rate on transcripts with long chains of thought rose from about 0% to about 3.8%.

A careful conclusion is narrower: Mythos Preview leads Claude Opus 4.6 on the key public benchmarks Anthropic selected, but public evidence does not establish that it is uniformly better for everyday writing, latency-sensitive chat, pricing, or every ordinary user workflow.

Can regular Claude users access it?

Not as a normal Claude model. Anthropic says Mythos Preview was not released for general availability; access was offered to partner organizations that maintain important software infrastructure, under terms restricting its use to cybersecurity.

So if your question is whether it appears as a switchable model in Claude.ai, whether there is standard API access, what it costs, how fast it is, or what its context window is, the public record does not provide a full product answer. Based on Anthropic’s disclosures, Mythos Preview is best described as a high-capability, high-sensitivity preview for defensive security work, not a new Claude tier for mainstream users.

Bottom line

Claude Mythos Preview publicly surfaced on April 7, 2026 with Project Glasswing, and Anthropic’s own benchmarks show it outperforming Claude Opus 4.6 across the listed evaluations. But Anthropic also says it has not made the model generally available and has limited access to cybersecurity-focused partners that maintain important software infrastructure.

For most readers, the practical answer is simple: Mythos Preview may be one of the most capable Claude models Anthropic has disclosed, but it is not currently a model ordinary users can simply choose and start using.