Vals AI projects Anthropic at researcher level in 2027
Vals AI's co-founder told The Information models may meet human-level researchers by August 2027 if the release pace holds. Forecast, not a lab deadline.
The evaluator’s co-founder told The Information that the current release pace points to August 2027. That date is Vals AI’s own extrapolation, not a deadline from Anthropic or OpenAI.
On September 29, 2026, The Information aired a TITV interview with Rayan Krishnan, CEO and co-founder of Vals AI. The San Francisco firm sells independent tests of AI models. The same day, the outlet posted a clip on X: if the past year’s release pace continues, “we’ll see models start to meet human level researchers by August 2027.” Krishnan added that OpenAI’s timeline for what Vals calls RSI “is a little bit longer,” and that Anthropic would get there first. OpenAI has not endorsed that comparison.
First: what Vals AI means by RSI
RSI, for recursive self-improvement, is the industry label for a model that helps build its successor: training, compression, tuning, and the tests around those jobs. Think of a junior engineer who already writes the experiment code. The Vals index is narrower. It scores four research and engineering tasks, run with a fixed budget, against a human or frontier reference. The analogy stops at judgment. The test does not measure a system that picks, on its own, what is worth inventing.
The scale runs from 0 (the starting baseline) to 1 (a theoretical ceiling), with 0.5 as a published reference. The index averages compression, language-model training, harness engineering (the tests and tools that hold an experiment together), and post-training, the extra tuning after the main training run. Vals treats 0.6 as a substantial gain, not full RSI. In the essay “When AI builds itself”, Anthropic wrote that it is not there yet and that recursive self-improvement “is not inevitable.”
What Krishnan said
Krishnan repeated the August 2027 mark in a post of his own on X. On the Vals leaderboard, read on October 8, 2026, the Claude family’s trend line points to July 2027 for frontier-researcher level. One month of difference is rounding on the same forecast. Vals puts Claude’s historical pace at about 0.30 index points a year. If that pace slows, the date slips.
In August 2026 Vals announced a $40 million Series A, at a $400 million valuation, led by Andreessen Horowitz (a16z). A company that sells the test has a stake in the index mattering. The scores below are according to Vals.
Who leads the index, according to Vals
Claude Opus 5.5 sits at 37.31% (0.373) and Claude Fable 5.1 at 36.09% (0.361). Google’s Gemini 4 Argon is next, at 30.55% (0.305). On the OpenAI side, GPT-6 Sol scores 28.1% (0.281) and GPT-6 Astra 27.1% (0.271). Vals says Opus 5.5 leads compression, language-model training, and post-training; GPT-6 Sol leads harness engineering. A headline that is only Anthropic versus OpenAI leaves Google out.
What Anthropic already admits
In the same essay, Anthropic says that as of May 2026 more than 80% of production code merged was attributed to Claude. In the second quarter of 2026, a typical engineer merged 8 times as much code per day as in 2024. The company itself warns that lines of code overstate productivity. The bottleneck, it says, is still judgment: the model can run a well-specified experiment; choosing the goal is still human work.
On September 11, 2026, CNBC reported public comments from alignment researchers at Anthropic and OpenAI, including Evan Hubinger, Anna Wang, and Jasmine Wang, about self-improvement speeding up before governance catches up. That coverage does not confirm Vals’s calendar.
What the forecast does not prove
August 2027 has not happened. It is not a commitment from Anthropic. OpenAI has not confirmed it is behind on this metric. There is no guarantee Claude will build the next model on its own. According to the interview recap, Krishnan also pointed to physical bottlenecks, chips and power, against an uncontrolled jump even if the index nears human level. We still do not know if the 0.30-point yearly pace will hold, or if 0.6 on this index would mean a model that chooses its own research goals.
Claude models lead a Vals test of research tasks. The missing piece is the part where the model decides what to research.
Sources
Press
- The Information, TITV: Evaluating OpenAI and Anthropic’s Latest Models
- The Information on X
- CNBC, Arjun Kharpal, Sept. 11, 2026
- Dealroom, recap of the TITV interview
Company (according to the source)