As of 2026-12-31, what is the ratio of the AI R&D uplift in value at Anthropic to the AI R&D uplift in value at OpenAI?

rationale:

The estimated ratio of Anthropic's AI R&D uplift in value to OpenAI's has a median of 1.05, reflecting a narrow edge for Anthropic in a race where both labs are likely experiencing comparable, modest overall acceleration. Crucially, "uplift in value" is a latent metric measuring the rate of valuable research output, not raw coding throughput. While both labs boast dramatic input metrics—Anthropic shipping 8x as much code, and OpenAI logging 3.1 agent-workdays per researcher workday Research acceleration: The view inside OpenAI—these inputs compress heavily when converted to actual research progress. For example, an 8x increase in code volume may only imply a 2.5x per-researcher uplift, which in turn might yield less than a 2x overall R&D acceleration due to compute bottlenecks and experiment cycle times Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x.

Both labs remain constrained by the need for human judgment in high-level research tasks, likely keeping their absolute value multipliers in the 1.3x to 2x range. At Anthropic, 0% of measured AI R&D work is fully autonomous (Level 5) Measurements for understanding the pace of AI .... Similarly, at OpenAI, the judgment-heavy Decide and Design phases—such as setting agendas and planning experiments—still account for less than 3% of total agent output tokens Research acceleration: The view inside OpenAI. Because both developers rely on top-tier internal models that are roughly at parity, and both face similar bottlenecks in human-led agenda setting, their multipliers are positively correlated, anchoring the ratio near 1.0.

Anthropic holds a slight estimated lead primarily due to deeper organizational adoption and the absence of recent internal friction. Claude currently "leads" (Level 4) roughly 26% of Anthropic's AI R&D work Measurements for understanding the pace of AI ... Measurements for understanding the pace of AI ..., and the company benefits from a preliminary third-party estimate by a METR team with elevated access, which suggested a roughly 1.5x overall acceleration with a 30% chance of reaching 2x Summary of METR's predeployment evaluation of Claude ... https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf. Additionally, CEO Dario Amodei has stated that AI progress is advancing "drastically faster" inside the lab https://darioamodei.com/post/we-must-pace-the-frontier. While lab testimony must be weighed carefully against hard evidence, Anthropic's consistent reporting and external validation paint a picture of highly integrated, relatively unencumbered AI assistance.

OpenAI, while scaling agent usage aggressively, has recently faced structural friction that likely drags on its value multiplier. Following a security incident in July 2026, OpenAI instituted a two-week reinforcement learning pause and moved Astra-class models into hardened, restricted environments https://openai.com/index/pacing-model-development-cyber-capabilities/ https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/. While OpenAI possesses a likely compute advantage over Anthropic—which generally eases the bottleneck between coding throughput and actual research progress—these post-incident security controls introduce procedural friction. Forecasts for OpenAI's revenue scaling strengthen the case that its massive compute scale could help it close Anthropic's efficiency gap, while the broader trajectory of open-ended agentic capabilities could allow that scale to translate into more valuable research output. Furthermore, OpenAI's own internal evaluations determined that GPT-6 Astra does not cross its "High" threshold for AI self-improvement.

The uncertainty intervals account for the fact that neither lab explicitly measures or publishes its "uplift in value," making this a comparison of two uncertain, latent variables. The ratio is estimated at 0.85 at the 10th percentile, 0.95 at the 25th, 1.05 at the median, 1.15 at the 75th, and 1.3 at the 90th. These somewhat narrowed extremes reflect the broader picture that shared compute bottlenecks will keep top labs tightly clustered. Much of Anthropic's apparent lead may stem from disclosure asymmetries, as METR's elevated-access estimate currently exists only for Anthropic Anthropic gives METR employee-like access after the ..., while OpenAI publishes entirely different input metrics. The ratio could shift below 1.0 if OpenAI's compute advantage, revenue-supported scaling, and rapid agent scaling overcome its security friction, or push higher if Anthropic deploys a new internal model generation that decisively cracks open-ended reasoning tasks before the end of the year.