AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

METR’s November 5 assessment said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, but emphasized high uncertainty and described primarily AI assistance to researchers. A separate Google DeepMind example reported that AlphaEvolve sped up one component of Gemini’s training by 23%, reducing total training time by 1%. These developments illustrate how AI could help improve future AI systems, while leaving open how much of research can be automated and how quickly gains will continue.

Independent evaluator METR said on November 5 that Anthropic is likely improving its models’ capabilities on well-specified tasks at around twice its earlier rate, while stressing that the estimate is highly uncertain. The assessment describes primarily AI assistance to researchers, not a fully autonomous system building its own successor. It adds evidence to a debate about whether AI tools could speed up the research that produces more capable AI.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

METR said it “would not be surprised” by a rate as low as 1.3 times the earlier pace or as high as three times. That range signals substantial uncertainty around the central estimate. METR also said Anthropic had redacted much of the evidence it considered most informative, limiting what readers can independently inspect from the assessment.

A separate example comes from Google DeepMind’s AlphaEvolve, which the company reported on May 14, 2025. DeepMind said the system improved a matrix-multiplication component used in Gemini by 23%, cutting Gemini’s overall training time by 1%. The component-level gain did not translate into a comparable reduction in total training time; the company’s reported result was a 1% overall reduction.

These examples concern different systems and measures. METR assessed the pace of model capability improvement on well-specified tasks at Anthropic, while DeepMind reported a specific engineering optimization. Neither example by itself establishes that AI research is fully automated or that model development will accelerate at a fixed rate.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentMETR published an assessment of AI-assisted research at Anthropic, reporting likely faster capability improvement with high uncertainty, as a Google DeepMind optimization example shows a smaller, concrete efficiency gain.
Graphic titled Around twice the earlier rate, with a wide range. METR, November 5, 2026: pace of capability gains at Anthropic on well-specified tasks compared with the historical rate of March 2024 to October 2025, as of October 2026. Highlighted: METR's overall judgment, about 2x, and it would not be surprised by 1.3x to 3x. Benchmark trend AECI since October 2025: 1.9x, 95 percent confidence interval 1.5x to 2.8x. Notes: R&D is accelerated primarily via AI assistance rather than fully autonomous contributions; Anthropic redacted a majority of the evidence METR found most informative for these conclusions
METR's finding and its range. Source: METR, "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026", November 5, 2026. Graphic: Thorsten Meyer AI.

How Research Gains Could Compound

The potential significance lies in a feedback loop: AI systems assist with research, researchers test the suggestions, and improvements that hold up may contribute to a stronger model. That model could then help with later research. If this cycle becomes reliable across more tasks, it could change how quickly laboratories develop models and how businesses plan around AI.

Efficiency improvements can also build on one another. In a simplified illustration, three independent savings of 10% each would leave a job requiring about 73% of its original compute. With a fixed compute budget, that would support roughly 37% more jobs of the same kind. This is a mathematical illustration, not a measured result from either lab, and real savings may not be independent or directly transferable.

Even incremental improvements could matter at scale if they reduce the time or computing resources needed for repeated experiments. But the practical effects would also depend on who can use the systems, who controls the infrastructure, and how institutions distribute the resulting gains.

Products Worth Considering

Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)

Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Feedback

AI research involves more than proposing a promising change. Researchers must inspect prior work and code, design experiments, run them under comparable conditions, investigate failures, repeat promising results, and decide whether the evidence supports adopting a change. A system that can assist with some of those steps may save time, but a reported improvement still needs fair testing.

For example, an apparent gain can result from an easier test, different compute use, or a result that disappears with a new random seed. Reliable research assistance must help identify such problems as well as generate ideas. The source essay presents a sequence in which a model assists with research, tests filter candidate changes, and improvements that survive are used in a stronger model. It describes this as a schematic of an argument, not as a measured process at any laboratory.

The cited evidence is limited in scope: METR’s assessment concerns one company’s pace of improvement on well-specified tasks, and DeepMind’s report concerns one component of Gemini’s training process. Together they offer examples of assistance and optimization, rather than proof that the full research cycle is already self-sustaining.

“would not be surprised”

— METR

Limits of the Current Evidence

METR characterized its estimate as highly uncertain, and the redaction of much of its most informative evidence makes independent scrutiny harder. The assessment also does not, from the information provided, establish how much of Anthropic’s research process AI handled, how much researcher supervision was required, or whether the estimated pace will persist.

The DeepMind result is a company-reported example of a specific optimization. The material does not establish whether similar gains can be repeated across other training components or research tasks. More broadly, experiment time, available hardware, data quality, continuity between research efforts, and the choice of research direction can all constrain progress. The pace may be uneven, and the sources do not settle how benefits or infrastructure access will be distributed.

Evidence Needed on the Research Loop

The next useful evidence would clarify how much research work AI systems complete, how much human oversight it takes, and whether gains hold up under repeated, comparable tests. Further assessments could also show whether results extend beyond well-specified tasks to broader research problems.

For now, the cited reports support cautious attention to AI-assisted research and engineering efficiency. They do not establish that a rapid, self-sustaining cycle is underway. Future measurements will need to show whether the assistance produces durable improvements across successive development cycles.

Key Questions

What did METR report about Anthropic?

METR said Anthropic is likely improving model capabilities on well-specified tasks at around twice its earlier rate, with high uncertainty. It said it would not be surprised by a rate between about 1.3 and three times the earlier pace.

Does the assessment show that AI can build a successor model by itself?

No. The assessment described work that was primarily AI assistance to researchers. The source material does not establish that AI independently completed the full process of designing, testing, and building a successor.

What did AlphaEvolve improve?

Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%.

Why could small efficiency gains matter?

If separate savings can be repeated across many jobs, a fixed compute budget could support more experiments or training work. The three-savings example in the source is an illustration; it is not a result measured by METR or Google DeepMind.

Source: Thorsten Meyer AI

You May Also Like

Auto Mode Is Now The Default In Claude Code

Claude Code has transitioned to auto mode as the default setting, enhancing user experience and automation capabilities.

AI output review queue for customer support macros

Support teams are testing a new AI output review queue for customer support macros to ensure policy compliance and tone accuracy before publication.

What xAI’s Grok Build CLI Actually Sends To xAI

Details emerge on the data transmitted by xAI’s Grok Build CLI, raising questions about privacy and data handling in AI development tools.

Wie Teuer Ist Es, Eine Selbstgehostete KI Im Vergleich Zu Forge?

Analyse der Kosten für selbstgehostete KI im Vergleich zu Forge, inklusive aktueller Preisentwicklung und Einfluss auf Souveränitätsentscheidungen.