Anthropic’s Leaked Mythos Model Claims to Outperform Opus 4.6 – Why Release Is Delayed

On March 26, an unpublished Anthropic internal blog draft was accidentally exposed when stored in an unencrypted, publicly searchable location. The draft, captured and mirrored by Twitter user @M1Astra, disclosed details about Anthropic’s next flagship model, referred to as Claude Mythos and internally codenamed Capybara .

The document positions Mythos as a "step‑change in frontier AI capability" that surpasses the current top‑tier Claude Opus 4.6. The claimed improvements focus on three areas:

Software programming : significantly better performance on code‑related tasks than Opus 4.6.

Academic reasoning : notable gains on complex reasoning and academic benchmarks.

Cybersecurity : highlighted as the most concerning and closely watched capability.

Anthropic unusually acknowledged the leak, confirming they are testing the model. Fortune reported the company’s admission, noting the model represents an "unprecedented cybersecurity risk" and that Anthropic intends to understand these near‑term threats before public release.

The draft outlines a cautious release strategy: early‑access users will initially be cybersecurity practitioners, allowing defenders to explore the model’s limits before potential attackers can exploit it. This approach aims to help secure codebases ahead of an "AI‑driven attack wave."

Another factor delaying a broader rollout is the model’s high compute cost. Anthropic admits that Mythos currently incurs extreme operational expenses, requiring efficiency improvements before it can be offered to general users.

According to Anthropic’s typical cadence, internal testing to full launch usually spans several months. Given the heightened focus on safety, the timeline may be even more conservative this time.

The final public name may differ from "Mythos"; the draft’s internal codename Capybara could change before release.