Claude Mythos is the most powerful AI model ever built. It is also the one you cannot use. Anthropic announced it on April 7, 2026, confirmed it is real, published a 244-page system card documenting what it can do, and then refused to release it to the public. That decision is unprecedented in the history of frontier AI development. No major lab has ever done this before.
This post covers everything confirmed about Claude Mythos: what it can do, why Anthropic will not release it, who has access right now, and what the realistic release date looks like based on Anthropic's own roadmap.
What Is Claude Mythos
Claude Mythos Preview is Anthropic's most capable model, sitting a full capability tier above the entire Claude Opus family. Anthropic calls this tier "Capybara" internally, introduced specifically because Mythos capabilities are qualitatively different from Opus-class models, not just quantitatively better.
The model is estimated at 10 trillion parameters using a Mixture of Experts architecture, meaning not all parameters are active during inference. Benchmark scores are genuinely unusual. Mythos scores 93.9% on SWE-bench Verified, 82.0% on Terminal-Bench 2.0, and 97.6% on USAMO 2026, which is a graduate-level mathematics competition. On long-context reasoning, graduate-level science, and the hardest agentic coding benchmarks, it does not just outperform competitors. It operates in a different category.
The UK AI Security Institute tested Mythos against a 32-step cyber range called "The Last Ones." Mythos solved it end-to-end 3 out of 10 times. It was the first AI model ever to do so. The next best performers were Claude Opus 4.6, followed by GPT-5.4 and GPT-5.3 Codex in a tie.
What Claude Mythos Can Do That Scared Anthropic
The cybersecurity capability is the reason Mythos is not public. The numbers are specific enough to take seriously.
During testing, Mythos autonomously identified thousands of zero-day vulnerabilities across every major operating system and browser. It found a 27-year-old bug in OpenBSD that had survived decades of expert human review without detection. It found a 17-year-old vulnerability in FreeBSD's NFS implementation, now tracked as CVE-2026-4747. These are not theoretical findings. These are real vulnerabilities in production systems used by millions of people.
Palo Alto Networks used Mythos during a testing period and found 75 vulnerabilities. In a normal month, their team would find five to ten. CrowdStrike's 2026 Global Threat Report documented an 89% year-over-year increase in AI-driven attacks, a figure their CTO used to argue the industry has no choice but to embrace AI-native defense.
In May 2026, a security firm used Mythos to crack macOS kernel protections within six days, bypassing Apple's Memory Integrity Enforcement — the hardware-backed security layer built for the M5 chip that Apple spent five years and billions of dollars developing. Apple has not yet shipped a fix. Mozilla used Mythos and found and patched 271 security vulnerabilities in Firefox in a single testing period.
Anthropic's own internal assessment warns that Mythos "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." That is not a vague safety concern. It is a specific warning about the asymmetry between offensive and defensive AI capability.
Who Has Access Right Now
Access to Claude Mythos runs through Project Glasswing, an invitation-only consortium of approximately 50 vetted organizations. Anthropic committed $100 million in model usage credits to the program, focused entirely on defensive cybersecurity work. Glasswing partner pricing is $25 per million input tokens and $125 per million output tokens, five times the cost of Opus 4.7.
The confirmed access channels for Glasswing partners are the Claude API, Amazon Bedrock research preview, Google Cloud Vertex AI private preview, and Microsoft Foundry. Standard API accounts cannot see or access the model identifier. If you try to call the Mythos model string from a normal Anthropic account, it does not exist.
The organization list is NDA-bound and has not been fully published. Public reporting names a mix of large cloud providers, security firms, defense-adjacent organizations, and a small number of independent safety evaluators. The NSA has reportedly used Mythos despite its parent organization, the Department of Defense, having blacklisted Anthropic after a dispute over autonomous weapons restrictions. China requested access and was refused.
The Google Cloud Label Change Explained
The story that triggered this week's speculation is the disappearance of the "Preview" label from Claude Mythos on the Google Cloud Vertex AI console. Developers who track AI releases noticed that every recent Anthropic model has followed the same sequence before going public: preview label removed, public blog post, general release. Claude Opus 4.7 followed this exact pattern.
Anthropic's own language contradicts the public release interpretation. The Project Glasswing page states clearly that they do not plan to make Claude Mythos Preview generally available. The system card states that the model's large increase in capabilities led to the decision not to make it generally available. The label change most likely reflects an internal staging update for existing Glasswing partners, not preparation for a public launch.
What the Actual Claude Mythos Release Date Looks Like
Anthropic has described a specific pathway toward eventually releasing Mythos-class capabilities to the public. The plan is to test and refine safety rails on less capable models first, then apply those safeguards to Mythos-class models once they are proven at scale.
Claude Opus 4.7, released April 16, 2026, is the first vehicle for this approach. It was deliberately trained with lower cybersecurity capabilities than Mythos and ships with automatic detection and blocking of prohibited uses. Anthropic has described Opus 4.7 as the first model in a sequence designed to develop the safety infrastructure needed before any Mythos-class release becomes possible.
Zvi Mowshowitz, who tracks AI development closely, has speculated that a Mythos-class model could reach general availability by September 2026. That is speculation, not an Anthropic commitment. A more conservative reading of Anthropic's own roadmap suggests late 2026 at the earliest, with 2027 being the more realistic window for a version of Mythos-class capabilities that meets their safety threshold for public release.
What Anthropic has confirmed is that their eventual goal is to enable users to safely deploy Mythos-class models at scale. The word "eventually" is doing a lot of work in that sentence.
Why This Matters Beyond the AI Community
The reason this story matters beyond the AI community is what it signals about the trajectory of AI capability development. The models available today through standard APIs, including Opus 4.7 which is already exceptional for most use cases, are below the threshold that caused Anthropic to pause.
For anyone building AI-powered systems, whether for trading, content automation, or financial analysis, the gap between what exists and what is publicly accessible has never been wider. The models coming in the next one to two years will be Mythos-class. The institutions and developers who understand how that capability lands, and what guardrails come with it, will be positioned very differently from those who do not.
Project Glasswing is the first serious answer to the question of how transformative AI capability gets deployed responsibly. The 50 organizations with access right now are essentially writing the playbook for what comes next. How that playbook gets applied to the public version is the story worth watching between now and whenever the Claude Mythos release date actually arrives.
There is also a financial angle that most coverage misses. The organizations inside Glasswing right now are building security products, compliance tools, and AI-native defense infrastructure using capabilities the rest of the market cannot access. By the time Mythos-class models reach general availability, those organizations will have a 12 to 18 month head start. In the AI economy, that is not a minor advantage. It is a structural moat. The companies that figure out how to use these capabilities first will not be competing with the companies that figure it out second. They will be operating in a different market entirely.
The same dynamic played out with early Bitcoin adoption and early institutional crypto access. The people who understood what the technology could do before the headlines caught up did not just make money. They built the infrastructure everyone else now depends on. AI capability is following the same curve. The window where Mythos-class capability is restricted is not a delay. It is a filter. And the filter is running right now.
Sources: Anthropic, Google Cloud, Cointelegraph
Disclosure: This post contains affiliate links to Kraken and Trezor. BitBrainers may earn a commission at no extra cost to you. This is not financial advice.
BitBrainers. We check the facts so you don't have to.
