TL;DR
METR’s November 5 assessment estimates, with high uncertainty, that Anthropic is improving model capabilities on well-specified tasks at about twice its earlier rate, mainly through AI assistance to researchers. Google DeepMind has separately reported that AlphaEvolve sped up a matrix multiplication component used in Gemini, cutting overall training time by 1%.
METR’s November 5 assessment says Anthropic is likely improving AI capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The assessment points mainly to AI assistance for researchers, offering evidence that AI tools may help accelerate the work of building future models.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
METR, an independent evaluation organization, said it “would not be surprised” if the rate were as low as 1.3 times or as high as three times the earlier rate. It characterized the work it described as primarily AI assistance to researchers. The estimate concerns capability improvements on well-specified tasks; it does not establish that AI systems are independently conducting all or most research.
METR also said Anthropic redacted much of the evidence it considered most informative. That limits what outsiders can assess about the underlying work and the basis for the estimate. The assessment is therefore an uncertain estimate, not a direct measurement of a fully automated research process.
A separate example comes from Google DeepMind, which reported in May 2025 that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. The company said this reduced Gemini’s overall training time by 1%. The result shows how an improvement to one component can affect a larger system, but it does not by itself demonstrate a broad or self-sustaining cycle of AI research.
How Research Gains Could Compound
If AI systems help researchers find, implement and test improvements, those improvements may feed into models that assist with later research. That creates a potential feedback loop: a model contributes to work on its successor, and the successor may become a more capable research aid. METR’s estimate makes that possibility relevant to current planning, while its wide uncertainty means the pace and scale of any effect remain unsettled.
Efficiency improvements can also matter when they apply repeatedly. A simple illustration in the source material assumes three independent 10% savings in the same job. Together, those savings would leave the job using about 73% of its original computing resources, allowing a fixed budget to run roughly 37% more such jobs. This is a hypothetical calculation, not a measured result from a laboratory. Real improvements may overlap, fail to generalize or require added costs.
The practical consequences would depend on more than technical capability. Access to advanced systems, ownership of the computing infrastructure and decisions about how gains are shared would shape who benefits. A faster research cycle could affect businesses and workers, but the cited evidence does not establish a particular economic outcome or timeline.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Bose SoundLink Max Portable Bluetooth Speaker, IP67 Waterproof and Dustproof, Up to 20 Hours of Battery Life, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
Evidence From Two Research Efforts
AI research involves more than generating ideas. Researchers must inspect prior work and existing code, design experiments, run them under comparable conditions, investigate failures and decide whether apparent gains hold up. A change that seems to improve performance may instead reflect an easier test, different computing resources or random variation. Reliable evaluation is needed to distinguish real gains from misleading results.
The source essay describes a possible loop in which a model proposes changes, helps write code and run experiments, and assists in checking results. Tests filter out ideas that do not hold up; improvements that survive may be incorporated into a stronger model, which then joins later research. The essay labels this sequence a schematic of its argument, not a measured process at any lab.
The two cited developments are distinct. METR assessed AI-assisted capability improvement at Anthropic in November 2026. Google DeepMind’s AlphaEvolve result, announced in May 2025, concerned a specific optimization to a component used in Gemini. Together they provide examples of AI assistance and system optimization, but they do not establish that every stage of the proposed research loop is already automated.
““would not be surprised” by a rate of 1.3x or 3x”
— METR
What the Estimate Cannot Establish
METR’s estimate has high uncertainty, and the organization said Anthropic redacted much of the evidence it found most informative. The published material therefore leaves outsiders with limited ability to examine the underlying evidence or judge how representative the described work is.
It is also unclear how much of AI research a system can complete reliably without close human supervision. The source identifies possible constraints including experiment time and hardware, the gap between small experiments and large systems, selecting research directions, continuity and data quality. These are considerations in the essay’s account; the cited assessment does not quantify how strongly each one limits progress.
The available examples do not show that an AI system can independently build a successor model or that gains will continue at a fixed rate. The timing, reach and distribution of any broader effects remain unknown.
What Future Evaluations Can Show
Further assessments can clarify whether AI assistance continues to raise the pace of capability improvement, and how much of the work models perform compared with human researchers. More evidence about methods and results would also help independent evaluators assess whether reported gains hold up across tasks and repeat trials.
For now, the available reports point to AI assistance in research and a specific training-efficiency improvement, alongside substantial uncertainty about their broader significance. How quickly the proposed loop develops will depend on technical limits as well as the choices organizations make about access and deployment.
Key Questions
What did METR estimate about Anthropic?
METR estimated that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate. It described the estimate as highly uncertain, saying it would not be surprised by a rate of 1.3x or 3x.
Does the assessment show that AI is doing all AI research?
No. METR characterized the work it described as primarily AI assistance to researchers. The assessment does not establish that AI systems are independently conducting all or most research.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%.
What remains uncertain about recursive AI research?
The scale and durability of the gains, how much human supervision is needed, and how quickly systems might take on more research tasks remain unclear. METR also said key evidence was redacted.
Source: Thorsten Meyer AI
