Anthropic's researchers noticed a change in their workflow after bringing AI into their daily routines. METR's assessment captured staff describing the shift as similar to gaining the help of postdocs and graduate students. One employee said they were now "only bottlenecked on ideas & compute, no longer bottlenecked on code." Those using Anthropic's top internal model reported about 12 times the productivity compared to working without AI, roughly double the rate seen with the previous system, Mythos 5, just months earlier. These numbers came from self-reports and haven't been independently checked across the company.
AI's Role in Accelerating Research
On November 5, 2026, the nonprofit Model Evaluation and Threat Research (METR) released an independent review of Anthropic's AI research and development speedup. The review drew on about 11 weeks of internal data and staff interviews. METR estimated that by October 2026, Anthropic's pace on "well-specified" AI tasks was about twice as fast as it had been from March 2024 to October 2025. This estimate used Anthropic's version of the Epoch Capabilities Index (AECI), which showed benchmark scores rising at an estimated 1.9 times the earlier rate, with a 95% confidence interval from 1.5 to 2.8 times. METR translated this into roughly two years of historical benchmark progress packed into the previous year. The group noted that the real acceleration could plausibly range from 1.3 to 3 times the earlier pace, showing there is still a lot of uncertainty.
The review centered on a feedback loop: researchers use AI to build better AI, which then helps further development. METR found most gains came from AI supporting human researchers, not from AI working alone. Staff interviews and surveys described AI being used to run experiments that used to take days and to try ideas that would have been too costly or slow before. There were early hints of more independent AI contributions in some areas, but fully autonomous research was not yet driving progress. METR said "the majority of the acceleration as of October is due to AI, with its contribution growing over time."
Evidence and Uncertainty
METR's analysis drew on public benchmark results, confidential internal research data, and interviews from mid-August through late October 2026. Anthropic shared internal information, but key parts of the evidence were not made public, which limits outside review of how benchmark trends, productivity data, and causal judgments were combined. METR presented its findings as an assessment of acceleration, not as proof that AI assistance alone caused the gains. Its public estimates included a central figure and wide uncertainty bounds, with METR saying it was "highly uncertain" and "would not be surprised by 1.3x–3x" acceleration compared to the earlier period.
To track progress, METR used the AECI to put results from different capability tests on a single scale. Since October 2025, AECI scores reportedly climbed at 1.9 times the pre-October-2025 rate. METR said it was "confident" the acceleration was not just from scaling up the computing power used to train Anthropic's models. This separated AI-driven acceleration from gains due to more training compute, which has long been a major factor in AI progress. The result is an estimated change in the rate of benchmark progress, not a direct claim that every Anthropic research activity became twice as productive. METR pointed out that "acceleration in AECI may not map cleanly to acceleration in real-world impact."
Attribution and Limitations
Figuring out how much of the acceleration came from AI itself was difficult. METR found that hiring and other factors likely played a role, especially in late 2025. By October 2026, the group judged that AI probably explained most of the observed acceleration, though this view relied partly on interviews and staff surveys. METR treated informal productivity polls with caution, noting that self-reported gains might not reflect actual improvements across the workforce. Employees described the AI support as similar to having "the support of postdocs and graduate students."
The report also listed reasons the acceleration estimate could be too high or too low. Structured tests may miss performance on less predictable tasks, and Anthropic's earlier progress was slower than the wider industry, which could make the acceleration look bigger than it is. On the other hand, benchmark results may lag behind recent changes, and faster growth in training compute did not explain the speed-up. The review's conclusion was strongest for well-defined tasks, where progress can be measured consistently, and weaker for open-ended scientific research, where quality and originality are harder to measure. METR limited its main claim to "well-specified tasks," rather than saying all research had doubled in speed.
Implications for AI Literacy and Education
For educators, students, and researchers focused on AI literacy, the Anthropic case shows both the promise and the complexity of using AI to speed up research. AI can help turn ideas into experiments faster and support new approaches, but the real effect on learning, teaching, or broader educational outcomes is still unclear. The findings point to the need for critical evaluation, transparent evidence, and careful attribution when judging claims about AI-driven progress in any field.
METR said the public must rely on its judgment more than it would like, given how much evidence remains redacted. This highlights a key challenge for AI literacy: understanding not just what AI can do, but how its contributions are measured, attributed, and communicated in research and education. The assessment came out as public debate continued over whether AI development should slow down, even as top labs use AI to speed up their own research. The findings matter beyond Anthropic, describing a possible self-reinforcing cycle: better models help research, faster research builds stronger models, and those models may further boost research productivity. METR's evidence did not show that Anthropic had reached fully autonomous AI research, since human researchers still handled much of the work. The current record shows rapid AI-assisted acceleration with uncertain causal attribution, not confirmed autonomous replacement of human research teams. For more details, see the METR independent assessment and Associated Press coverage.
