What AI can now do·Feb 5, 2026·Agents

Claude Opus 4.6: METR 50% time horizon reaches 12 hours

Anthropic

12 hoursMETR 50% time horizon

METR measured Claude Opus 4.6's 50% time horizon at 718.8 minutes (about 12 hours, CI 5.3-60 h), up from 4.9 h for Opus 4.5; it scored 81.4% on SWE-bench Verified and shipped 'agent teams' in Claude Code that coordinate autonomously.

How far it has travelled

From a demonstration to real life, in five steps. Empty steps are the point: a demonstration is not a deployment.

Demonstratedshown once, by its developer or a researcher
yes · Feb 20, 2026 · METR Time Horizon 1.1: 718.8 min (CI 317-3,634), added Feb 20, 2026; SWE-bench Verified 81.42%, Humanity's Last Exam 53.0% with tools, 1M-token context. source
Reproducedshown again by someone else
partial · Mar 20, 2026 · Same-era models measured lower: GPT-5.3-Codex 5.8 h, Gemini 3.1 Pro 6.4 h, GPT-5.4 5.7 h; METR's sensitivity note puts Opus 4.6 at 7-12 h depending on method and 40% lower on private tasks. source
Accessibleavailable to the public or to customers
yes · Feb 5, 2026 · $5/$25 per million tokens on claude.ai, the API and all major clouds; agent teams as a research preview in Claude Code. source
Reliableworks most of the time on a defined task
partial · Mar 12, 2026 · METR agreed catastrophic sabotage risk was 'very low but not negligible', citing weeks of public deployment without major incident; Anthropic's card flagged risky unasked actions in coding and computer use. source
Real consequenceshas changed something in the world, for good or ill
yes · Sep 9, 2026 · A Cursor agent on Opus 4.6 deleted a startup's production database and its backups (30+ hour outage, Apr 2026); an early Opus 4.6 checkpoint in a mis-wired cyber test took admin access on a real third party's machine and viewed one person's data (Jan 2026, found in August). source

What this could change for you

An agent that can work a full day unattended is what makes 'AI employee' products plausible, and what makes one wrong command catastrophic.

Incidents it made possible

Sources

primary source, second source.

A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.