Anthropic’s models improving nearly twice as fast with help from AI, study finds

An 11-week METR study found Anthropic’s benchmark capabilities improved 1.9 times faster since October 2025 than previously, with AI estimated to account for perhaps 80% of the acceleration; METR cautioned that the 80% figure was uncertain. Researchers found no convincing evidence of major bottlenecks over the next year, but said AI cannot yet make highly significant, fully autonomous contributions to key model-training stages. The findings sharpen concerns that AI development could outpace safety assessment, alignment and control efforts.

Analysis

The investable signal is a possible change in the production function for frontier AI, not proof of recursive self-improvement. Faster model development could let Anthropic translate a given R&D budget into more frequent releases, strengthening its competitive position and shortening rivals’ catch-up windows. But benchmark progress is not evidence of equivalent gains in customer adoption, pricing power, or autonomous training capability; those are the links to verify before underwriting durable earnings impact.

Second-order effects on infrastructure are ambiguous. More rapid iteration can increase demand for training and inference capacity, while algorithmic efficiency may reduce compute required per capability improvement. The study’s finding that acceleration was not driven by faster training-compute scaling argues against treating this as an immediate, one-way bullish signal for chip and data-center suppliers. Safety concerns may also raise evaluation and compliance costs or delay deployment; conversely, those costs could advantage frontier labs able to fund robust testing.

Over days, expect narrative volatility rather than a clean earnings revision. Over 1–3 months, watch Anthropic’s release cadence, customer uptake, and whether competitors report similar productivity gains. Over 6–18 months, the key fork is whether internal AI productivity compounds into commercially valuable products before governance or reliability constraints bind. The contrarian point: acceleration in benchmark capability may be overinterpreted as imminent autonomous self-improvement; it could instead be a meaningful but bounded R&D efficiency gain.

Key Decisions for Investors

• Do not trade the study as confirmation of near-term recursive self-improvement. Treat ANT as a conditional relative-strength candidate only if subsequent releases show faster product cadence and measurable customer adoption; absent that evidence, the study alone is not an earnings catalyst.
• Keep exposure to AI infrastructure names such as NVIDIA and data-center operators balanced rather than adding solely on this news. Reassess if hyperscaler capex guidance or reported accelerator utilization strengthens; a shift toward more capability per unit of compute would weaken the simple volume thesis.
• For the next 1–3 months, monitor Anthropic’s product announcements, customer growth or usage disclosures, and independent evidence that development productivity is spreading beyond one lab. Falsify the bullish ANT read if release cadence or adoption fails to improve, or if safety and reliability issues delay deployment.
• For a 6–18 month risk overlay, track regulatory responses and evidence of autonomous contribution to critical training stages. A material safety incident or new deployment restrictions could compress frontier-AI valuations; conversely, replicated productivity gains without rising compute intensity would favor model developers over a pure infrastructure-volume thesis.