Part of: this research — view all 8 rows
The 2× Anchor and Current Thresholds The clearest anchor for Anthropic's current AI R&D uplift is its own internal counterfactual estimate and subsequent threshold declarations. In its private August 2026 Risk Report (covering up to mid-July), Anthropic estimated its speedup relative to an unassisted counterfactual at "slightly less than a factor of 2," which METR calculated as roughly 1.97× Review of the "Risks from automated R&D" section in .... Since then, Anthropic has aggressively ramped up internal model use—introducing Opus 5.5, Fable 5.2, and stronger unreleased models—and determined on November 10 that it has met its 2× Responsible Scaling Policy threshold. Because this recent determination was explicitly based on expected progress from future models rather than strictly observed current outputs Anthropic, PBC - S-1/A-2 Anthropic's IPO filing says it has met its AI-speeds-up-AI line ..., it does not establish that current uplift has crossed 2.0×. I anchor current uplift at 1.95×, reflecting the consensus that observed progress is very close to but just below that threshold.
Capability Progress vs. Value Uplift METR’s independent assessment corroborates a figure near 2×, finding that Anthropic’s capability progress on well-specified tasks runs at about 1.9× its March 2024–October 2025 baseline Independent Assessment of AI R&D Acceleration .... METR attributes approximately 80% of this acceleration to AI assistance Independent Assessment of AI R&D Acceleration .... Converting benchmark progress to value uplift requires two key adjustments. First, the historical baseline already included some AI assistance; rebasing the ~1.7× AI-attributable capability gain against a strict no-AI counterfactual pushes the estimate back up toward 2×, though the uncertain conversion does not justify placing the central estimate above the observed-progress anchor of 1.95×. Second, while raw input metrics have exploded—Claude now writes over 90% of merged codebase changes and daily code volume per engineer is up 45×—these do not translate linearly to research value. Anthropic explicitly models strong diminishing returns, where a doubling of labor input yields only a 1.15–1.3× increase in research output Measurements for understanding the pace of AI ..., naturally constraining the true value uplift despite massive coding throughput.
Downside Risks and Misattribution There are meaningful reasons to suspect the true research uplift could be closer to the 1.4× range. METR’s capability index showed only a 1.4× acceleration when all developers' models were included Is Anthropic on the cusp of RSI? Inside the redacted METR ..., and Google DeepMind’s own recent tests found no significant end-to-end acceleration. This suggests some benchmark gains may reflect industry-wide trends, increased compute, or a one-time jump rather than sustained AI uplift Where's the “intelligence explosion”? - by Ramez Naam. Furthermore, METR’s assessment faced significant limitations: it lacked employee-like access and query access to non-public models, dropped compute verification from its scope, and noted substantial redactions alongside disagreement among Anthropic’s own staff regarding attribution Why a new report on Anthropic automating AI development ... Independent Assessment of AI R&D Acceleration .... These uncertainties put my 10th percentile at 1.45× and my 25th percentile at 1.7×.
Upside Potential and Latent Contributions Conversely, the true uplift in value could exceed the observed-progress anchor if current internal research is running significantly ahead of observed capability trends. METR explicitly notes a training-time lag between research breakthroughs and their appearance in benchmarked models Independent Assessment of AI R&D Acceleration .... Anthropic’s own reports of extensive pre-release research use and the existence of unreleased internal models scoring higher than Mythos 5.2 https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card suggest that latent productivity gains may not yet be visible to external evaluators. However, this upside is heavily constrained by Anthropic's own admission that its newest models do not yet fully substitute for senior researchers. I place the 75th percentile at 2.2× and the 90th percentile at 2.6×, allowing for latent contributions while keeping the median at 1.95×, just below the 2× threshold.
The current uplift was anchored at 1.95x, reflecting the consensus that observed progress is very close to but just below the 2x threshold.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited