Question
As of 2026-11-14, what share of the experiments that materially informed the most recent frontier model release at the largest AI lab by revenue were proposed by an AI system rather than by a human researcher?
Target Lab and Baseline: Under the August 2026 definitions, OpenAI remains the largest AI developer by revenue (though differing accounting practices make Anthropic's run-rate comparable 2 sources). This makes OpenAI's most recent frontier release the target—likely GPT-6 Cyber (November 5) or GPT-6.1 Terra (October 26). The experiments that materially informed these models ran primarily between May and October 2026. Across the frontier, AI is heavily utilized for executing research rather than proposing it. For example, METR's assessment of Anthropic found researchers leveraging AI to rapidly execute human ideas, describing the ability to "spam 10 ideas" at agents while remaining "bottlenecked on ideas & compute."
Tokens vs. Experiments: There are no publicly reported direct measures of AI experiment origination, necessitating reliance on proxies. OpenAI's internal metrics from August 2026 show that the "Decide" and "Design" phases combined accounted for just 2.8% of agent output tokens, with "research and experiment planning" at 0.7% and "what to work on" at 0.3%. However, these are shares of tokens, not of experiments by count. Because proposing an experiment requires orders of magnitude fewer tokens than implementing, coding, and debugging it, this token share serves as a soft floor rather than a direct translation or cap on the experiment-count share.
Adoption and Capabilities: The ceiling for AI origination is loosely bounded by OpenAI's report that experiments per active experimenter rose from 0.72× to 1.60× the 2025 average. Even if the entirety of this excess were attributed to AI-proposed ideas, the count share would peak around 37%—but realistically, most of this increase represents human ideas executed faster via coding agents. Capability evaluations do suggest that proposal quality is rising: next-step evaluations indicate models' proposed steps are judged better than a human researcher's 50% to 64% of the time 2 sources, and Anthropic notes that Mythos 5.2's advantage over Opus 5.5 stems mostly from judgment in choosing next steps anthropic.com. Yet, adoption lags behind capability. The State of AI Report notes that agents are mostly used "post-experimental ideation and design" stateof.ai, and OpenAI explicitly maintains that humans "still set our research priorities, judge which ideas and results to pursue."
The "Materially Informed" Filter and Uncertainty: The central estimate reflects that while high-level agenda setting is securely human, AI systems are increasingly originating local ablations, hyperparameter optimization points, and diagnostic follow-up runs within automated loops. The wide confidence interval is driven by definitional ambiguity. If the final resolution strictly filters for major, release-shaping novel hypotheses, the "materially informed" clause pulls the share down into the low single digits, as the most substantive architectural and data-mixture choices remain human-driven. Conversely, if every agent-selected sub-run or diagnostic sweep point within a human-assigned task is liberally counted as a distinct AI-proposed experiment, the true share pushes well into the double digits.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited