Part of: this research — view all 12 rows
As of mid-November 2026, frontier developers rely heavily on AI for implementation and experiment execution, but none has stated that a released model was principally designed by an AI. The scale of automation is massive: Claude authors over 90% of Anthropic's codebase contributions, Google DeepMind's production code is more than 75% AI-written, and OpenAI runs 3.14 agent-workdays per researcher-workday. METR’s recent assessment of Anthropic found progress on specific R&D tasks at around twice the historical rate, attributing roughly 80% of this acceleration to AI. Despite this, labs emphasize that humans remain firmly in control of research direction. OpenAI reports that people still set priorities and judge results openai.com, while METR notes that AI is not yet making highly significant autonomous contributions to the most important parts of model training metr.org.
The resolution criteria demand an explicit claim of principal design—such as architecture, training recipe, or data mix—rather than mere coding or tactical optimization. Current examples fall short of this standard: AlphaEvolve's documented Gemini contribution was a kernel optimization reducing training time by 1% deepmind.google, and older systems like NASNet operate within heavily constrained, human-defined search spaces research.google. However, underlying capabilities are advancing steeply. AI research taste evaluations (TasteVal) show a doubling time of about three months, and Anthropic recently stated that it expects to meet its Responsible Scaling Policy's AI R&D threshold based on the trajectory of future AI models.
A qualifying statement could emerge through two distinct pathways. First, a leading developer might demonstrate AI-led design on a smaller, specialized, or experimental released model, which would meet the criteria without requiring full researcher replacement. Second, the claim could come from a lower-ranked frontier developer—such as xAI or a top-ten Chinese lab—motivated by marketing incentives to promote recursive self-improvement. We have already seen a non-frontier startup claim in October 2026 that its automated system chose its model's architecture, quantization, and compute kernels on its own 2 sources. A similar explicit claim from a qualifying frontier developer is a plausible near-term trigger.
Conversely, major labs face strong safety, regulatory, and liability pressures that discourage framing AI models as 'principally designed' by AI. Anthropic's recent SEC filings explicitly anticipate the need for 'meaningful human supervision' and float potential policy updates requiring human review and approval of R&D decisions 3 sources. OpenAI also emphasizes keeping humans in the loop, targeting an automated researcher working 'under human direction' or 'in tandem.' These framing incentives mean that even if an AI originates most substantive choices, labs are highly likely to describe the design as human-led or human-approved, directly delaying or preventing a qualifying statement.
These competing dynamics center the median in late 2028, aligning with the expected maturation of automated researcher systems and the subsequent lag to release an AI-designed model. The early tail accounts for a potential hype-driven claim from a lower-tier frontier lab in mid-to-late 2027. Meanwhile, the wide upper tail extending well past the 2032 horizon reflects the significant possibility that strict human-in-the-loop mandates, pacing commitments, or government regulations permanently deter developers from officially attributing principal model design to an AI.
Evaluated alongside related research milestones, the timeline holds firm, tracking slightly behind experiment selection milestones as principal model design remains highly sensitive .
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited