How much of the work of building new AI is already done by AI at Anthropic and OpenAI, by their own count, and when people expect AI to build AI alone.
26%
Anthropic: share of its AI research work that Claude leads
1 Aug 2026. Anthropic Institute, Measurements for understanding the pace of AI development inside frontier labs
Tap a dot, or use the arrow keys, to see its number and source.
Anthropic says Claude now runs about a quarter of the work of building its next AI from start to finish, with a person checking rather than steering, up from almost none in February 2026.
64%
1 Apr 2026 (Claude Mythos Preview, labelled Apr 2026). Anthropic Institute, When AI builds itself
Tap a dot, or use the arrow keys, to see its number and source.
When an Anthropic researcher went down a wrong path, Anthropic's newest AI suggested a better next step 64% of the time, up from 22% for a 2024 model; this is the judgement part of research that AI has been weakest at.
3.1
15 Aug 2026 (OpenAI says 'as of mid-August' 2026; day not given). OpenAI, Research acceleration: The view inside OpenAI (2026-09-06)
OpenAI says its AI agents now put in about three days of work for every day its human researchers work, up from less than one day before June 2026.
Vals AI: Jul 2027
"Anthropic is on track for frontier-level AI researchers by July 2027."
Vals AI (RSI Index), 1 Oct 2026. Vals AI, Vals RSI Index (updated 10/1/2026)
Trend extrapolation, not a promise: Vals says Anthropic's RSI Index score has risen about 0.30 points a year and on that trend reaches 0.6 (results well beyond published human references on its research tasks) in July 2027. Vals first posted the July date on X on 2026-09-22 after testing Claude Opus 5.5; its earlier date was August 2027 (CEO Rayan Krishnan on X, 2026-09-23: 'We've updated our timeline for full RSI to July 2027 instead of August'). The index changed from five tasks to four between versions, so its scores are not shown as a time series. No source found for 'March 2027' from Vals.
OpenAI: Mar 2028
"We are making strong progress toward creating an automated AI researcher by March of 2028."
OpenAI, 6 Sep 2026. OpenAI, Research acceleration: The view inside OpenAI
A company goal, first announced in October 2025. In the same post OpenAI says it has met its earlier goal of an 'automated research intern' by September 2026, by its own measurements, and that it does 'not yet know how to safely get all the way to aligned, full RSI.'
Jack Clark: Dec 2028
"I think there's a ~60% chance we see automated AI R&D (where a frontier model is able to autonomously train a successor version of itself) by the end of 2028."
Jack Clark (Anthropic co-founder), Import AI, 4 May 2026. Jack Clark, Import AI 455: Automating AI research
Same essay: 'If you had to push me for a 2027 probability, I'd say 30%.'