As of 2026-10-16, 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?

rationale:

Lab and Release Context Based on recent disclosures, Anthropic is treated as the largest AI lab by revenue, having reached an annualized run rate exceeding $65B at the end of July 2026 https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html, compared to OpenAI's reported ~$50B in late September https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/ https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/. On this basis, the most recent frontier model release is Claude Opus 5.5, which shipped on September 22, 2026 https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf https://www.anthropic.com/news. The experiments that materially informed this model would have run roughly between May and September 2026, relying primarily on Mythos 5- and Opus 5-class internal assistants.

Implementation vs. Proposal This assessment strictly distinguishes between experiments proposed by an AI and those merely implemented by one. Headline automation metrics heavily overstate AI's role in experimental design. While over 80% of Anthropic's merged code is authored by Claude and the model "leads" (AL4) 26% of AI R&D work, these metrics evaluate task execution https://www.anthropic.com/institute/measuring-pace-of-ai-development https://www.anthropic.com/institute/recursive-self-improvement. Anthropic's automation scale explicitly does not rate who sets agendas or who proposes and selects experiments https://www.anthropic.com/institute/measuring-pace-of-ai-development https://www.anthropic.com/institute/measuring-pace-of-ai-development.

Current Limits on Agenda-Setting Evidence from the frontier indicates that high-level experimental design remains overwhelmingly human-driven. Anthropic observes that large performance gaps persist when it comes to Claude exercising judgment in choosing goals https://www.anthropic.com/institute/recursive-self-improvement https://www.anthropic.com/institute/recursive-self-improvement, a point reinforced by the Opus 5.5 system card noting the model lacks researcher "judgment" or "taste" https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf. Both Anthropic and OpenAI confirm that humans still set research priorities and judge which ideas and results to pursue https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf Research acceleration: The view inside OpenAI. Furthermore, OpenAI's internal proxy data shows that "research and experiment planning" accounts for just 0.7% of agent output tokens, while the broader "Decide plus Design" phase makes up only 2.8% Research acceleration: The view inside OpenAI Research acceleration: The view inside OpenAI. While token share is a floor-shaped proxy, it underscores that high-level planning is a minimal fraction of current AI R&D workloads.

Sub-Experiments and Autonomous Loops Conversely, there is clear momentum in localized experiment generation. Claude's success rate on the most open-ended internal tasks rose from 26% to 91% by September 2026 https://www.anthropic.com/institute/recursive-self-improvement https://www.anthropic.com/institute/recursive-self-improvement. In tasks where an agent "leads," it completes the workflow end-to-end from a high-level ask, which plausibly involves autonomously proposing diagnostic runs, hyperparameter sweeps, and ablations. Anthropic has also noted instances, such as its weak-to-strong researcher demonstration, where agents designed experiments themselves while humans only set directions https://www.anthropic.com/institute/recursive-self-improvement https://alignment.anthropic.com/2026/automated-w2s-researcher/. Counted strictly by number, these agent-driven sub-experiments constitute a growing share of total runs.

Conclusion The median estimate of 10% balances the reality of agent-driven sub-experiments against the requirement that these experiments materially informed the release, while weighing the broader evidence that few high-level research decisions are currently entrusted to agents. This filter naturally privileges larger, human-directed structural experiments over minor automated ablations. The central 50% range of 6% to 18% reflects uncertainty over how often autonomously proposed diagnostic runs, hyperparameter sweeps, and ablations materially affected the final model. The lower bound of 3% reflects a strict interpretation where only major, human-signed-off experiments clear the threshold. The upper bound of 28% accommodates a more generous definition, assuming that the sheer volume of diagnostic runs and hyperparameter sweeps autonomously proposed by Claude during AL4 tasks represents a substantial fraction of the total experiment count behind Opus 5.5.