Website owners and marketers can use Anthropic’s accelerating AI development to plan for faster changes in search visibility, content production and AI-generated recommendations. METR’s high-access evaluation from August 12 to October 29, 2026 estimates that Anthropic was improving model capabilities at about twice its March 2024 to October 2025 rate by October 2026.
The finding is based on benchmark growth, internal research data, staff interviews and surveys. Anthropic redacted much of the evidence METR considered most informative, so the public must rely heavily on METR’s judgment when evaluating the result.
What Did METR Find About Anthropic’s AI Progress?
METR said Anthropic’s benchmark scores, measured through AECI, Anthropic’s version of the Epoch Capabilities Index, grew at an estimated 1.9 times the pre-October 2025 rate. The 95% confidence interval was 1.5 to 2.8 times, which means Anthropic appeared to complete roughly two years of historical benchmark progress in one year.
That estimate falls between the development of Claude 2.1 and Sonnet 4.5. METR also found that the acceleration was not caused by faster growth in training compute. Its robustness checks produced acceleration estimates ranging from 1.1 to 2.9 times the historical rate.
Most METR staff involved in the assessment believe Anthropic’s improvement rate on well-specified tasks was about twice its earlier rate. METR said the actual rate could be closer to 1.3 or three times the historical rate. It also reported significant uncertainty about how the gain translates to broader tasks that are less precisely defined.
Biomedical Research Acceleration Study
Anthropic’s AI research acceleration has a parallel in the 2025 Biomedical Research Acceleration Study, which examines how AI may affect research productivity and scientific discovery. The study identified accelerating scientific discovery as a potentially profound impact of AI, while emphasizing that narrow AI had already been used in biomedical research, including AlphaFold protein structure prediction.
The 2025 Biomedical Research Acceleration Study placed general-purpose AI within a progression from narrow systems toward broader capabilities. It looked at the wider role of AI in research and its limits, rather than evaluating acceleration inside an AI company. For readers assessing Anthropic’s development pace, that gives broader context for the potential value of AI-assisted research without changing the separate limits of Anthropic’s evaluation.
- Biomedical Research Acceleration Study examined AI, research productivity and scientific discovery in 2025.
- Narrow AI was already part of biomedical research through applications such as AlphaFold protein structure prediction.
- General-purpose AI represented a progression toward broader capabilities.
How Much of the Acceleration Came From AI?
METR’s subjective estimate is that about 80% of the October 2026 acceleration came from Anthropic’s use of AI in research and development, although the assessment gives this estimate substantial uncertainty. Anthropic staff interviews, surveys, self-reported productivity gains, experiment usage, code output and headcount data all pointed toward AI assistance as an important factor.
The evidence suggests that AI was assisting researchers rather than independently producing the work. METR also saw signs that fully autonomous contributions could become more important soon.
Non-AI factors likely influenced the acceleration observed in late 2025. METR investigated compute scaling, hiring, improved data, algorithmic discoveries and the resolution of other bottlenecks. It found convincing evidence that faster training-compute growth was not responsible for the recent increase, but it could not decisively explain the remainder of the earlier acceleration.
What Does This Mean for Website Search Visibility?
Faster model development can change the way people discover websites and the kinds of information search systems can retrieve. The arXiv paper What if automating AI R&D triggers an intelligence explosion?, led by Alan Chan with 21 other authors, says AI systems already write most of the code inside the companies building them. It also concludes that AI systems are on track to automate most AI research and development work within a few years, and possibly all of it.
For website owners, the practical effect is a shorter planning cycle. Improvements in automated research, code generation and content systems may help teams produce and update pages faster. The same pace can increase competition for visibility, so clear information, structured pages and accurate business listings remain important as discovery systems change.
The paper also emphasizes that faster capability growth could compress years of advances into months and create risks that society cannot easily manage. The assessment by METR says Anthropic’s researchers had not found a clear bottleneck that would prevent further progress over the following 12 months, but its understanding changed as more information became available.
What Would a Site Audit Flag?
A conventional website audit will not directly measure Anthropic’s internal research acceleration. It can, however, reveal whether a site gives automated discovery systems accurate, accessible information that can support search visibility.
SEOScanPro’s technical audit checks the measurable issues behind each on-page finding, which can help a site owner identify problems such as inaccessible content or other technical constraints. For broader AI visibility, BizScoreAI can examine how a business appears in AI search and test how an AI agent interprets a website. These tools address the website and listing signals under a site’s control, while METR’s findings describe the accelerating technology environment in which those signals must perform.
The linked paper Forecasting AI Time Horizon Under Compute Slowdowns, written by Parker Whitfill, Ben Snodin and Joel Becker, also adds caution to forecasting agent capabilities. Its model treats AI time-horizon growth as proportional to compute growth. Under one projection using OpenAI’s compute trajectory, a one-month task horizon at 80% reliability would arrive seven years later than a simple trend extrapolation would suggest.
What to Do Now
- Review the information that automated systems can read on priority pages, including headings, product details, location information and consistent business data.
- Run a technical audit to identify measurable page issues that could prevent crawlers or AI agents from reading and interpreting the site.
- Check how the business appears when an AI agent searches for its products or services, then correct any inaccurate or incomplete business listing.
- Plan content and site improvements in shorter cycles because research automation and AI-assisted development may increase the pace of change.
- Separate measured site findings from wider AI forecasts, since METR’s acceleration estimate has substantial uncertainty and does not directly describe how search systems rank a particular website.
What Are the Main Uncertainties?
METR cautioned that it could underestimate acceleration if recent growth rates have increased further, AI research tasks become more automated, or AECI has understated real-world effects. It could also overestimate acceleration if non-AI factors contributed more than expected, AECI overstates progress on broadly useful tasks, or Anthropic’s 2023 to 2025 growth was unusually low.
Anthropic’s redactions were the largest limitation. METR said some redacted evidence had multiple plausible interpretations, and both METR and Anthropic researchers had substantial uncertainty about how to combine it into a complete picture.
FAQ
How Fast Was Anthropic Improving Its AI Models in October 2026?
METR estimated that Anthropic was improving model capabilities on well-specified tasks at about twice its March 2024 to October 2025 rate. METR said the true rate could be closer to 1.3 or three times the historical rate.
What Benchmark Did METR Use to Assess Anthropic’s Progress?
METR used AECI, Anthropic’s version of the Epoch Capabilities Index. It estimated that benchmark scores had accelerated by 1.9 times since October 2025, with a 95% confidence interval of 1.5 to 2.8 times.
Did Anthropic’s Faster Progress Come From More Training Compute?
METR said it received convincing evidence that the acceleration was not caused by an acceleration in the scale of training compute. Details supporting that conclusion were redacted.
Try the AI visibility report
SEOScanPro runs a full technical audit of a site and shows the measured result behind every check. Open the AI visibility report.