Question
As of 2026-11-14, by what percentage does adopting each successive frontier model release raise research productivity at the largest AI lab by revenue, per unit of Epoch Capabilities Index improvement?
Target Lab and the Metric Based on recent realized revenue of $33.2 billion for the first three quarters of 2026 and an annualized run rate near $81 billion, Anthropic is currently the largest frontier AI developer by revenue Anthropic, PBC - S-1/A-2 OpenAI seeks to accelerate fundraising ahead of Anthropic .... The parameter to estimate is the marginal increase in valuable research output per unit of research effort—not just raw coding volume. While AI wrote over 90% of merged lines at Anthropic in late 2026, compute and experiment cycles sit between coding throughput and overall R&D progress Anthropic, PBC - S-1/A-2 Independent Assessment of AI R&D Acceleration .... Because of these bottlenecks, the effective research productivity in value is significantly lower than intermediate automation metrics imply.
Observed Progress vs. Capability Gains To price the elasticity per ECI point, we anchor on the observed acceleration in capability progress and the pace of the capability frontier. An independent assessment by METR found that Anthropic’s progress rate increased by approximately 2× (with a 95% interval of 1.3× to 3×) since October 2025, attributing about 80% of this to AI Independent Assessment of AI R&D Acceleration .... Anthropic’s internal fit indicates an acceleration of 1.88× Claude Fable 5.2 & Claude Mythos 5.2 System Card. I use 1.95× as the observed capability progress acceleration anchoring the central estimate . Over roughly the same period, Epoch's fitted trend for reasoning-era models shows the capability frontier advancing by about 16.6 ECI points per year Updating the ECI methodology to use a Bayesian model.
Translating Progress to Productivity Because sustaining progress requires growing inputs against diminishing returns in research, a 2× acceleration in the progress rate requires a larger increase in effective research productivity, plausibly an effective labor uplift of 2.5× to 5× https://www.lesswrong.com/posts/Jga7PHMzfZf4fbdyo/if-mythos-actually-made-anthropic-employees-4x-more Anthropic's AI acceleration is probably real - Beyond Imitation. Over a span of roughly 16 to 25 ECI points from the start of significant AI coding assistance to current frontier models, this supports a central estimate of 4.5% per point, calibrated to the observed 1.95× capability progress acceleration . For scale, compounding a 4.5% per-point increase over 16.6 ECI points yields about 2.1× annual growth in research productivity. This is a consistency check on the scale of the estimate, not an equality between productivity growth and progress acceleration: the translation also depends on diminishing returns, growing research inputs, and the period over which assistance accumulated.
Evaluating the 15% Threshold These estimates sit well below the 15%-per-point threshold required for self-sustaining acceleration. If the true value were 15% per point, research productivity would compound by approximately 10× per year at the current 16.6-point ECI pace. This is heavily contradicted by Anthropic's own determination that its newest released models do not cross its R&D threshold individually Claude Fable 5.2 & Claude Mythos 5.2 System Card and by Google DeepMind's finding of no statistically significant acceleration in their own tests Google Says AI Isn't Building Itself Yet—Even Though It .... Furthermore, Anthropic’s recent company-level threshold determination was explicitly based on expectations for future models rather than the observed performance of current models Anthropic, PBC - S-1/A-2 Anthropic Met RSP AI R&D Threshold in S-1/A-2.
Uncertainties and Tail Risks The right skew in the distribution accounts for the possibility that the productivity uplift is highly convex near current frontier capabilities, meaning the marginal effect of recent capabilities might be higher than the historical average. We must also weigh that Anthropic redacted much of the evidence METR found most informative Independent Assessment of AI R&D Acceleration ..., and unreleased internal models sit above the public frontier, which could mean the latent parameter is somewhat higher today. Conversely, the lower tail recognizes that a significant portion of recent acceleration could be driven by tooling, scaffolding, or increased inference compute rather than pure underlying model capability gains. My median is therefore 4.5% per ECI point, with the 25th and 75th percentiles at 2.8% and 7.5%, and the 10th and 90th percentiles at 1.5% and 12.0%.
The estimate was tightened around 4.5% per capability index point to stay mathematically consistent with the observed 1.95x capability progress acceleration .
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited