Breaking

Vals AI Puts Anthropic Ahead of OpenAI on AI That Researches Itself, With a July 2027 Mark

September 29, 2026 at 18:02 EDT

Foundation Models
AI Agents
Research & Papers

Anthropic’s latest Claude models lead an independent test of whether an AI system can do the research that builds the next model, and the benchmark’s authors project that Anthropic reaches a frontier-level research score by July 2027 if the past year’s pace holds. Vals AI co-founder and chief executive Rayan Krishnan has said he thinks OpenAI’s timeline for recursive self-improvement is a little longer. Neither company says a model is yet designing its own successor without people.

September 29, 2026 · Frontier labs

Claude leads the test of AI that researches AI

If the past year’s pace holds, Anthropic is on track for a frontier-level research score by July 2027. OpenAI’s own target for an automated researcher is March 2028. Neither lab says a model is building its successor alone.

Who leads autonomous AI research

Vals RSI Index. Column height tracks the score. Other labs are shown at their stated ceiling, 28.1%.

37.31%

Claude Opus 5.5

36.09%

Claude Fable 5.1

33.02%

Claude Opus 5

28.1%

Other labs, at most

Takeaway: three Claude models hold the top scores. GPT-6 Sol still wins harness engineering, one of the four tasks.

Code each engineer merges per day

Anthropic, second quarter of 2026 versus 2024. One block equals the 2024 pace.

2024

Q2 2026

Takeaway: by May 2026, Claude authored more than 80% of code merged into Anthropic’s codebase, up from the low single digits before February 2025. People still direct and review it.

Where the lead is real

Claude holds the top three index scores and is the first model under this protocol to beat a published language-model training reference. Inside Anthropic, coding output per engineer is far above 2024.

Where it still falls short

Agents mostly recombined known methods and were weak at choosing better research bets. GPT-6 Sol leads harness engineering. Full autonomy is unclaimed, and OpenAI says it does not yet know how to reach safe, aligned RSI.

What is automated, and what is not

1 · Set a task

A fixed research problem and a compute budget

→

2 · Run experiments

The agent iterates inside a standard harness

→

3 · Submit one result

Graded against a reference, not a hunch

→

4 · People still decide

The step back to a new frontier model is not closed

What 0.6 does not mean

July 2027 is a research score, not a model training its replacement alone.

A 0.5 matches selected references. A 0.6 is a larger gain, on the order of historic method shifts, still inside a human-set budget. OpenAI’s March 2028 goal is an automated researcher working with staff. Both labs say fully autonomous recursive self-improvement is not here.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.