🔍 Read the full analysis: How The AI Tool Could Help Build Its Successor on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
METR reported on Nov. 5, 2026, that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, while stressing substantial uncertainty. The work it described was primarily AI assistance to researchers, not autonomous AI research; a separate Google DeepMind example shows how targeted efficiency gains can reduce training costs.
An independent assessment by METR, published Nov. 5, found that Anthropic was likely improving its models’ capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. METR said the work it examined was primarily AI assistance to researchers—a development that matters because models helping with experiments, code and analysis could contribute to building more capable successors.
METR said it would not be surprised if the acceleration were as low as 1.3 times the earlier rate or as high as three times. Those figures describe METR’s uncertain assessment, not a directly measured, settled rate. The organization reviewed internal information from August through October 2026; the public account does not provide all of the underlying evidence.
According to the source account, Anthropic redacted much of the material METR considered most informative. That limits outsiders’ ability to independently evaluate the assessment. METR’s conclusion was that AI-assisted research likely explained most of the acceleration it observed, but the work described was mainly researchers using AI, not systems independently running the full research process.
A separate example comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. That is a specific reported engineering improvement, not evidence by itself that AI systems can autonomously develop new models.
How Research Gains Could Compound
The potential significance lies in a feedback loop: researchers use AI to improve parts of model development, and those improvements may produce systems that can help with more of the next development cycle. If that assistance is reliable, a research team could investigate more ideas, write and test code faster, or spend less time diagnosing routine failures. A stronger successor might then provide more capable assistance. The METR assessment offers evidence consistent with an early version of that process, while leaving its scale uncertain.
Efficiency gains can also add up. In a simplified illustration, three independent 10% reductions in compute leave a job using about 73% of its original compute. Under the same fixed compute budget, that would allow about 37% more of those jobs. This is an illustration, not a measurement of actual AI research productivity: real savings may not be independent, and additional experiments still take time and require suitable hardware and people to assess results.
For businesses and the public, the pace matters because it affects how quickly capabilities, costs and demand for work might change. But technical progress alone does not settle who can use powerful systems, who controls the computing infrastructure, or how benefits are distributed. Those outcomes depend on deployment choices and institutions as well as research performance.
Products Worth Considering

Bose Smart Ultra Dolby Atmos Soundbar

Apple Watch Series 10 GPS 46mm with Silver Aluminum Case and Blue Cloud Sport Loop

Samsung Galaxy Tab S10 Lite, 10.9-inch LCD, 6GB RAM, 128GB Storage, Gray
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than proposing a new model design. A team might review previous work, inspect an implementation, prepare data, write code, configure an experiment, compare results and repeat promising runs. Each step offers opportunities for assistance, but also ways to produce misleading evidence. A reported improvement could result from an easier test, unequal compute budgets or random variation rather than a genuine advance.
For an AI contribution to be useful, researchers need to check whether a result holds up under fair comparisons and repeat testing. A system that proposes an experiment or changes code is not necessarily able to judge whether its result is meaningful. The source article argues that progress should be measured by how much of this sequence systems can handle, how much supervision they need, and whether their contributions survive independent checks.
There are practical limits to the feedback loop. Experiments take time and hardware; results from small tests may not carry over to large systems. Researchers still need to choose worthwhile directions, maintain continuity across work and judge data quality. The source account describes progress as potentially uneven, rather than a smooth or guaranteed acceleration.
What the Assessment Cannot Establish
The public account does not disclose the full evidence METR reviewed, because Anthropic redacted much of the material the evaluator considered most informative. The estimate therefore cannot be independently reconstructed from the details provided here. METR’s range also signals that the size of the acceleration remains uncertain.
It is not clear how much of the reported improvement came from AI assistance specifically, which research tasks benefited most, or whether gains will persist as systems and experiments become more demanding. The assessment concerns well-specified tasks and primarily assisted work; it does not establish that AI can independently select research goals, validate findings or develop a successor model end to end.
The AlphaEvolve example likewise concerns a particular component and a reported training-time reduction. It does not show that comparable savings are available across all model development, or that engineering efficiencies will translate directly into a faster overall pace of capability growth.
Evidence to Watch in Future Evaluations
The next useful evidence would include further independent evaluations that explain their methods and distinguish AI assistance from work completed autonomously. Readers will also need clearer information about which tasks were performed, how results were checked, and whether reported gains held up across repeated experiments and larger systems.
Until then, the available material supports a measured conclusion: AI tools may be helping researchers improve models more quickly, but the size and durability of that effect are not settled. The path toward systems that can build substantial parts of their own successors remains a possibility, not a confirmed outcome. How quickly it develops will depend on research reliability, available compute, human oversight and decisions about deployment.
Key Questions
What did METR report about Anthropic?
METR said Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It emphasized high uncertainty, saying it would not be surprised by an estimate from 1.3 to three times the earlier rate.
Did METR find that AI was building models on its own?
No. The source account says the work METR described was primarily AI assistance to researchers. The assessment does not establish that AI systems independently handle the full process of designing, testing and validating successor models.
What did AlphaEvolve change in Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%, according to the source account.
Why is the METR estimate uncertain?
METR’s estimate included a wide range, and Anthropic redacted much of the evidence METR considered most informative. The public account does not provide enough detail for outsiders to independently reproduce the assessment.
What would show that AI research acceleration is continuing?
Further independent evaluations could help by reporting methods, separating assisted from autonomous work, and showing whether improvements survive repeated tests and transfer to larger systems.
Source: ThorstenMeyerAI.com
