Claude Mythos Preview reaches 17-hour horizon

The headline number is impressive, but the bigger story is that Anthropic’s restricted model is now brushing against the ceiling of METR’s current eval methodology, so the exact 17-hour figure should be treated as directional, not precise.

• –This is a benchmark result, not a public launch, and it comes from METR’s updated time-horizon page rather than a fresh product release
• –The task suite is weighted toward software engineering, ML, and cybersecurity, so the score says more about agentic technical work than general-purpose autonomy
• –Being above the 16-hour reliability limit means small changes in tasks or scaffolding could move the estimate noticeably
• –Even with that caveat, the result reinforces the same theme as Anthropic’s security-focused framing: frontier models are getting better at sustained, tool-using work
• –For developers, the practical takeaway is that long-context and multi-step agent evals are becoming a real differentiator, not just a lab curiosity