On 2028-06-30, which of these will best describe frontier AI research practice at the largest AI lab by revenue?

rationale:

Which Lab Sets the Standard Determining the largest AI lab by revenue currently centers on Anthropic and OpenAI. Anthropic reported an annualized run rate above $65B in July 2026, while recent estimates place OpenAI around $50B https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/. Because both labs describe remarkably similar internal practices, their competition for the top revenue spot matters little to the resolution. Both clearly operate under a baseline where humans direct and AI implements, meaning a complete stagnation ("No material change from 2026" at 2%) is highly unlikely given the steep 22-month trajectory ahead.

The Rapid Automation of Implementation The execution of research tasks is being automated rapidly. By August 2026, Claude was "leading" (AL4) 26% of Anthropic's AI R&D work, with its success rate on the most open-ended internal tasks climbing from 26% to 91% over a short window https://www.anthropic.com/institute/recursive-self-improvement https://www.anthropic.com/institute/recursive-self-improvement. OpenAI similarly reported a steep climb in agent use, formally declaring they had reached an "automated research intern" capable of multi-day tasks under human direction Research acceleration: The view inside OpenAI. This momentum, together with the stronger expected automation of intermediate tasks implied by timelines for major R&D automation claims, makes it almost certain that practice will progress well beyond simple implementation.

The Stickiness of Agenda-Setting Despite these execution gains, high-level planning and agenda-setting remain firmly human. OpenAI's metrics show that planning and design tasks ("Decide" plus "Design") constituted only 2.8% of agent output tokens in August 2026, with "what to work on" making up just 0.3% Research acceleration: The view inside OpenAI. Anthropic explicitly notes that "large performance gaps persist" in Claude's goal-choosing judgment https://www.anthropic.com/institute/recursive-self-improvement. Because agenda-setting is the most guarded function and is restricted by safety frameworks like the Preparedness Framework and the RSP, "AI sets research agendas with humans auditing outcomes" remains a distinct tail risk at 10%.

The Mid-2028 Core Modalities The vast majority of probability mass is split between the two intermediate states. "AI proposes experiments and humans select" (32%) perfectly maps to the labs' own current phrasing that humans "judge which ideas and results to pursue" Research acceleration: The view inside OpenAI https://www.anthropic.com/institute/recursive-self-improvement. As agent design capabilities improve, presenting options for human selection is the most natural workflow, especially because massive compute costs and alignment risks—such as OpenAI's July 2026 RL pause—demand human approval before resources are committed. Slightly more likely is "AI runs the full experiment loop while humans set research agendas" (38%), reflecting stronger expected automation of intermediate tasks when set against timelines for major R&D automation claims and aligning directly with OpenAI's stated goal of deploying an "automated AI researcher" by March 2028 while humans remain "in charge and setting the goals" OpenAI is throwing everything into building a fully ....

Factors Keeping Human Direction Primary There remains an 18% chance that the dominant practice is still best described as "Humans direct and AI implements." Progress in agent reasoning could stall, scaling returns might flatten, or compute constraints could bottleneck full experiment loops. Additionally, the political climate surrounding the "pace the frontier" movement strongly incentivizes lab leadership to publicly frame humans as firmly in control of direction and implementation. A major safety incident involving rogue agents could also trigger tighter internal limits, halting the progression of autonomy and keeping AI strictly in an implementation role.