METR’s evaluation, published on November 5, shows that Anthropic employees informally reported an approximately 12x productivity boost from using its latest models, while the pace of improvement in its model capability index was about 1.9x that of the past. The figures deserve attention, but they cannot be added together or treated as an increase in R&D output.
What this is
METR is a nonprofit organization that evaluates whether frontier models can complete real-world R&D tasks. The 12x figure in this report came from an internal Anthropic Slack poll in which employees used emoji reactions to describe how much the latest models had improved their work. It indicates that the models are more helpful in some parts of the job, but respondents may not have fully estimated how their working time changed, and this was not a randomized controlled experiment.
The 1.9x figure comes from METR’s AECI capability index. The report treats October 2025 as an inflection point: in the earlier period, the index increased by roughly 15 points per year, compared with about 29 points annually afterward. The main estimate is 1.9x, with a 95% confidence interval of 1.5x to 2.8x. Depending on the benchmark, date range, or treatment of saturation, the estimated range is approximately 1.1x to 2.9x.
A faster rate of index improvement does not mean that the model’s total score is now 1.9x its previous level, much less that every task is 1.9x faster. Whether faster individual work translates into team results also depends on code review, queues, failed experiments, and time spent waiting for collaborators.
Industry view
We believe the important signal is this: model capabilities are improving faster on tasks that can be clearly defined and verified, which will continue to reduce the time cost of some research workflows. But executives should not put “employees feel faster” directly into return-on-investment projections, nor should they treat a benchmark’s slope as organizational productivity.
The risks are equally clear. METR says the current acceleration may reflect changes in hiring, data improvements, algorithmic discoveries, and other bottlenecks, and it estimates that roughly 80% comes from AI. That is a researcher attribution, however, not a causal share identified through a randomized experiment. Benchmarks may also be affected by overfitting to their distributions, metric design, and how representative their tasks are. The report also argues that current progress comes mainly from AI assistance, rather than from models becoming capable of independently producing major scientific contributions.
Impact on regular people
For enterprise IT: The more defensible approach to procurement and deployment is to target measurable process changes, such as coding assistance, document retrieval, and test generation, rather than promise a blanket 12x efficiency gain. Pilots should track cycle time, delivery quality, and rework rates simultaneously.
For individual professionals: AI will most likely compress repetitive work first, giving people who understand models and can break work into tasks a larger advantage. Whether teams actually produce more still depends on how clearly requirements are defined and how effectively people collaborate.
For consumer markets: Capability gains will gradually show up in software responsiveness and the cost of producing content. In the near term, however, the common outcome will be faster performance in individual steps—not every organization achieving a 12x productivity boost at once.