🔍 Read the full analysis: The Next AI May Be Built With Help From AI on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
An independent METR assessment published Nov. 5 judged that Anthropic’s AI capabilities on well-specified tasks are improving at about twice their earlier rate, with substantial uncertainty. METR said AI assistance to researchers likely accounts for most of the acceleration it assessed; the public evidence is limited by redactions, and the finding does not establish that AI systems are independently building their successors.
An assessment published Nov. 5 by independent evaluator METR found that Anthropic is likely improving its AI capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. METR attributed most of the acceleration it assessed to AI assistance for researchers, offering a concrete—but incomplete—indication that AI tools may help speed the development of future models.
METR’s assessment drew on information from August through October 2026. The evaluator 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. Its estimate concerns capability improvements on well-specified tasks; it is not a claim that all AI research, or model development overall, has doubled in speed.
The work METR described was primarily AI assistance to human researchers, rather than systems conducting the entire research process independently. Such assistance could include helping with code, experiments, or analysis. The source material does not establish how much of each research task models can complete, how often their contributions survive review, or how broadly the finding applies beyond Anthropic.
Public scrutiny is limited because Anthropic redacted much of the evidence METR considered most informative. Separately, Google DeepMind has reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. That is a company-reported example of a targeted efficiency improvement, not evidence by itself that AI systems are autonomously designing successor models.
Why Research-Speed Gains Matter
If AI tools help researchers test more ideas or diagnose routine failures faster, a laboratory could make progress with the same staff and resources—or direct those resources toward more experiments. The potential significance is not limited to the quality of one model: the research process itself may become more productive when the systems being developed can assist with their development.
Efficiency gains can also build on one another, though the effect depends on whether improvements are independent and apply to the same work. As a simplified illustration, three separate 10% reductions in computing needs would leave a task requiring about 73% of its original compute. Under that assumption, the same computing budget could run roughly 37% more such tasks. This is an illustration, not a measured result from the METR assessment or AlphaEvolve report.
A faster research cycle could affect how companies plan investment, staffing, and infrastructure, and how governments and communities prepare for changes in AI capability. But technical progress alone does not determine who can use these systems, who controls the computing infrastructure, or how gains are distributed. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering

Dyson Cyclone V10 Origin Cordless Vacuum

Samsung 55-Inch M80H Series 4K Mini LED Smart TV (2026 Model)

eufy C30 Robot Vacuum and Mop
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than proposing a promising idea. A team must inspect prior work and existing code, prepare data and experiments, measure results, investigate failures, and decide whether a result is reliable enough to use. An apparent gain can be misleading if a test became easier, runs used different computing budgets, or the result disappears under another random seed.
For that reason, useful automation requires more than producing code or suggesting an experiment. Researchers need to check whether a proposed change was tested fairly and whether the evidence supports the conclusion. Human review and independent checks remain relevant when a system’s contributions are uncertain or could introduce errors.
METR’s assessment makes the discussion more concrete by examining reported acceleration at one AI company. A separate example from Google DeepMind shows how a system can contribute to a specific efficiency improvement. Neither example, on the information provided, demonstrates a complete loop in which an AI independently chooses a research direction, validates its discoveries, and builds a stronger successor without substantial human involvement.
“It “would not be surprised” by a rate of 1.3x or 3x.”
— METR
Limits of the Public Evidence
The estimate carries substantial uncertainty, and the assessment’s public evidentiary basis is incomplete: Anthropic redacted much of the material METR considered most informative. The available account does not give enough detail to independently assess every underlying measurement or determine how representative the observed work is of AI research more broadly.
It is also unclear how much of the reported acceleration came from AI tools compared with other factors, how consistently AI contributions hold up under repeated testing, and whether gains on well-specified tasks translate into progress on less structured research problems. The examples do not establish when, or whether, systems will be capable of directing and validating an entire model-development effort. The pace could be constrained by experiment time, computing hardware, data quality, continuity of research, and the difficulty of choosing promising directions.
What Researchers Must Verify
The next useful evidence will show, in greater detail, what research tasks AI systems perform, how much supervision they need, and whether their results remain valuable after independent replication and review. Further assessments could clarify whether the apparent acceleration persists over time and how much it varies across tasks or organizations.
For now, the METR assessment is a qualified signal rather than proof of fully automated AI development. Researchers, companies, and policymakers will need to track both the reliability of AI-generated research contributions and the limits imposed by hardware, experiments, and institutional choices. The report does not specify a next publication date or milestone.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate. It gave a broad uncertainty range, saying it would not be surprised by a rate of 1.3 times or three times the earlier rate.
Does the assessment show that AI is building new models by itself?
No. The work METR described was primarily AI assistance to researchers. The assessment does not establish that an AI system independently runs the full process of designing, testing, and building a successor model.
Why is the finding uncertain?
METR expressed substantial uncertainty, and Anthropic redacted much of the evidence the evaluator considered most informative. That limits what outside readers can independently verify from the public account.
What does the AlphaEvolve example show?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. It is a reported efficiency improvement in a specific component, not proof of autonomous AI research.
Source: ThorstenMeyerAI.com
