The AI Model Anthropic Doesn't Want You to Use Yet
On March 26, 2026, a security researcher discovered something unusual: a misconfigured data store on Anthropic's own infrastructure was leaking thousands of internal files — publicly, without authentication. Among those files was a detailed draft blog post describing an AI model called Claude Mythos, internally codenamed Capybara.
What the documents revealed has set the AI industry talking: Anthropic has trained a model it describes as a "step change" beyond Claude Opus — and it is so capable, the company is deliberately withholding public access out of safety concerns.
What We Know About Claude Mythos
- Capability claim: Internal documents state the model is "currently far ahead of any other AI model in cyber capabilities"
- The warning: Mythos "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders"
- Parameters: Approximately 10 trillion — a significant leap over all current public models
- Current access: Limited to select cybersecurity partners only
Why Is Anthropic Holding It Back?
Anthropic is not withholding Claude Mythos for commercial reasons — they believe releasing it too widely, too fast, creates a genuine security risk. The staged release logic:
- Cybersecurity partners get access first to build defences
- Access expands to Claude Pro, Team, and Enterprise only once defensive tools are mature
- Public API access comes last — timeline "determined by safety evaluation outcomes, not a commercial schedule"
Claude Mythos vs Current Public Models
| Model | Status | Key Strength | Public? |
|---|---|---|---|
| Claude Mythos | Cyber partners only | Advanced cyber, 10T parameters | ❌ No |
| Claude Sonnet 4.6 | Live | Expert real-world work, coding | ✅ Yes |
| GPT-5.4 | Live | Autonomous tasks, 1M context | ✅ Yes |
| Gemini 3.1 Pro | Live | Benchmark leader, multimodal | ✅ Yes |
What This Means for the AI Industry
The Claude Mythos situation raises questions the industry will grapple with for years: Who decides when an AI is safe enough to release? Are other labs holding back similarly powerful models? Prediction markets currently give roughly 25% probability of a Mythos public launch by April 30, 2026 — with the majority favouring a June 30 release.
Frequently Asked Questions About Claude Mythos
Q: What is Claude Mythos?
Claude Mythos is Anthropic's next-generation AI model, internally codenamed Capybara, accidentally leaked in March 2026. It reportedly runs on 10 trillion parameters and is described as far ahead of any existing public AI in cybersecurity capabilities.
Q: When will Claude Mythos be publicly released?
No confirmed public date exists. Anthropic says the timeline depends on safety evaluations. Prediction markets point to a June 2026 release window as the most likely scenario.
Q: Is Claude Mythos more powerful than GPT-5.4?
Based on leaked documents, Anthropic internally positions Mythos as a "step change" beyond all current models. No public benchmark comparisons exist yet — but the 10 trillion parameter count dwarfs every currently available AI model.
Q: What is the most powerful AI model available right now in 2026?
Among publicly available models, Gemini 3.1 Pro leads on benchmarks and Claude Sonnet 4.6 leads on real expert work. Claude Mythos is more powerful but not yet public. See: Best AI Models of 2026 — Full Comparison.
What to Watch Next
If Anthropic expands Claude Mythos access before June, or if OpenAI responds with GPT-5.5 (codenamed "Spud" — already in late pretraining), the AI landscape could shift dramatically in the next 60 days.
Also read: GPT-5.4 Review 2026 — OpenAI's AI That Works Like a Digital Employee