Picking the right problem was the last thing the lab said it could not yet automate. This is the measurement of that gap closing: in cases where researchers hit a dead end, Claude improved on their result 64% of the time, up from 22% in 2024, and its success rate on open-ended problems rose 50 points to 76% in six months, at quality the lab rates on par with humans and projects to pass them within the year (steering; open-ended success). The result is the judgment half of the When AI builds itself report, complementing the 52x speedup on execution; it follows Jack Clark’s January observation that models were improving at components of research and anticipates the August finding that research taste can be trained by reinforcement learning.
Development
A model learns to abandon dead ends better than its researchers
Problem selection was the capability the lab itself named as the missing piece for runaway recursion, so a measured rise in the model's judgment about when to abandon a line of work is the gap-closing number in the report.
Claim
Claude improved on researchers who hit a dead end 64% of the time, up from 22% in 2024, and its success on open-ended problems jumped 50 points to 76% in six months at quality already on par with human and projected to pass it within the year.
Reported metric: 64% of the time, up from 22%
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to June 5, 2026 (theinnermostloop.substack.com)
- Anthropic on X: Claude improves on researchers who hit a dead end 64% of the time (x.com)
- Anthropic on X: success on open-ended problems up 50 points to 76% (x.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.