Part of: this research — view all 8 rows
The target is the ratio of Anthropic's AI R&D value-uplift multiplier on 2029-12-31 to its baseline on 2026-11-14. This multiplier measures the rate at which Anthropic produces research it judges valuable with AI assistance compared to without; it is not simply coding throughput. The current baseline is estimated at roughly 2×. METR's independent assessment of Anthropic's capability progress puts it at about twice the previous rate (with a plausible range of 1.3× to 3×) and attributes approximately 80% of this acceleration to AI Independent Assessment of AI R&D Acceleration .... Anthropic's own index fits show a 1.88× acceleration, though its November 2026 threshold determination relied on expected future capabilities rather than observed current models Anthropic, PBC - S-1/A-2. While coding metrics show Claude writing over 90% of merged lines, end-to-end research value incorporates bottlenecks like experiment cycles, keeping the true current baseline near 2×.
By late 2029, automation will likely encompass much more of the research pipeline, driving substantial multiplier growth. Claude's footprint is expanding rapidly: it went from "leading" under 1% of AI R&D work in February 2026 to 26% by August https://www.anthropic.com/institute/measuring-pace-of-ai-development. Anthropic's internal task success approaches 90%, and unreleased models already score higher than released ones on internal indices. Furthermore, METR finds no bottleneck for the next 12 months Independent Assessment of AI R&D Acceleration ... and cannot rule out substantial further acceleration within six months Review of the "Risks from automated R&D" section in .... Supported by about $900 billion in compute commitments, a central expectation (a median ratio of 4.5×) balances the compounding benefits of AI automation against physical and experimental bottlenecks. This reflects a scenario where AI agents reliably handle experiment design, diagnosis, and evaluation, substantially speeding up valuable research output to an absolute multiplier around 9×, without eliminating the constraints on realizing that value.
However, physical and organizational bottlenecks keep the lower tail grounded (p10 of 1.5, p25 of 2.5). Even as AI agent capabilities improve, the realization of research value is constrained by the need for experimental compute, serial training cycles, and human verification. Anthropic explicitly identifies review, direction-setting, compute, and infrastructure as limiting factors When AI builds itself. Experimental compute can absorb much of the apparent task-level speedup The Intelligence-Explosion Paper Rests on One Number, ..., as demonstrated when it accounted for 70–90% of agent-run costs in METR's experiments Expenditure Horizon: Measuring Optimization Ability, with .... Additionally, deliberate pacing could artificially limit growth. Anthropic is considering requiring human review of R&D decisions Anthropic, PBC - S-1/A-2, and its leadership has called for "speed limits" on recursive self-improvement https://darioamodei.com/post/we-must-pace-the-frontier. A strict internal policy, a plateau in capabilities, or external regulation could hold the ratio near 1.5, while continued automation gains that remain substantially constrained by these bottlenecks support a lower-quartile ratio of 2.5.
The right tail (p75 of 8.0, p90 of 18.0) remains heavily skewed to account for the possibility of rapid, software-driven recursive self-improvement, while allowing physical and experimental bottlenecks to persist even with extensive automation. If models acquire strong research judgment and automate validation entirely, today's human and experimental bottlenecks may become temporary or be bypassed through better optimization. In a scenario where AI handles nearly all research autonomously, the counterfactual "no-AI" baseline becomes difficult to define, and overall value output could explode. This near-full automation tail allows for the ratio to reach beyond 20×, translating to massive absolute gains in end-to-end research progress, but such outcomes sit beyond the 90th percentile rather than defining the central expectation.
The expected growth ratio was centered around 4.5x, balancing the compounding benefits of AI automation against physical and experimental bottlenecks.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited