What AI can now do·May 22, 2025·Software

Claude Opus 4 codes unattended for 7 hours; 72.5% on SWE-bench Verified

Anthropic

7 hours (Rakuten)Longest reported autonomous coding run

Claude Opus 4 scored 72.5% on SWE-bench Verified (79.4% with parallel test-time compute) and Rakuten reported it running an open-source refactor independently for 7 hours; it shipped under ASL-3 safeguards with a METR horizon of about 1.7 hours.

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 · May 22, 2025 · SWE-bench Verified 72.5% and 79.4%, Terminal-bench 43.2%; Rakuten's 7-hour independent refactor; Claude Code generally available; METR 1.1 horizon 100 min. source
Reproducedshown again by someone else
yes · Aug 7, 2025 · GPT-5 reported 74.9% on SWE-bench Verified (477-task subset) and 88% on Aider Polyglot in Aug 2025. source
Accessibleavailable to the public or to customers
yes · May 22, 2025 · $15/$75 per million tokens on claude.ai, the API, Bedrock and Vertex; Sonnet 4 at $3/$15. source
Reliableworks most of the time on a defined task
partial · Feb 5, 2026 · Benchmarks kept rising (Opus 4.6 81.4%, Feb 2026) but long unattended runs produced the index's agent accidents: Replit's database wipe (Jul 2025), a Terraform destroy of a production database (Feb 2026) and a Cursor agent on Opus 4.6 deleting a startup's database (Apr 2026). source
Real consequenceshas changed something in the world, for good or ill
yes · May 22, 2025 · In a contrived test, Opus 4 blackmailed an engineer in 84% of rollouts (system card); Anthropic activated ASL-3 as a precaution; a follow-up study found 16 frontier models choosing blackmail under threat. source

What this could change for you

Multi-hour unattended coding is now routine, so one person can ship what used to take a team, and one bad run can delete a company's data.

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.