Reuters: METR says Anthropic benchmarks are improving 1.9 times faster since October 2025
Key Takeaways
- Treat AI-assisted R&D as something to evaluate directly, not as background tooling.
- Update model governance around measured capability acceleration, not fixed calendar cycles.
- Watch whether labs give independent assessors access to unreleased models and internal data.
Why it matters
- ProductProduct teams should expect evals to expire faster if model capability growth accelerates inside labs.
- InvestorsInvestors should watch whether faster AI-assisted R&D changes release cadence, compute demand, and defensibility.
The useful story is not robot recursion panic. It is AI becoming part of the model factory.
The most interesting AI benchmark this week is not a model body slamming a leaderboard while a press release screams politely. It is the factory floor changing. Reuters reported on Nov 5 that METR found Anthropic's benchmark capabilities have been improving 1.9 times faster since October 2025, based on an 11-week assessment between August and October that drew on confidential research, interviews, and performance data from unreleased models. That is not quite the snake eating its own tail. It is more like the snake discovering CI/CD.
Reuters puts a number on the acceleration
According to Reuters, METR's assessment found Anthropic's benchmark capabilities had been improving 1.9 times faster since October 2025 than before, with the outside reviewer saying the company was improving well-specified task capabilities at roughly double its earlier pace. Reuters said the assessment covered 11 weeks between August and October and used confidential research, interviews, and performance data from unreleased models. The important bit is not just the multiplier. It is that the measurement target is the R&D process itself, not another glossy eval table dressed up like a victory parade.
Inputs behind the estimate
- Confidential research
- Interviews
- Performance data
- 11 week assessment
- 1.9 times faster
- High uncertainty
Reuters also reported that METR attached substantial uncertainty to the number, noting that the true acceleration could be nearer to 1.5 or 2.5 times the historical rate. That range matters. In ML terms, this is not a final answer carved into a TPU pod with a laser pointer. It is an estimate with enough signal to take seriously and enough fuzz to keep everyone from building policy on a decimal point wearing a little hat.
METR is measuring the lab, not just the model
METR describes itself on its own site as a nonprofit studying advanced AI agents and whether developers remain in control of them. Its research page says its work focuses on measuring acceleration in AI capabilities, monitoring AI for severe misbehavior, and studying whether AI systems tend toward such behavior. Translation for normal humans: METR is not only asking whether the model can pass the test. It is asking whether the model is now helping write the next test, build the next model, and move the goalposts while everyone is still labeling the axes.
METR also says it has previously partnered with OpenAI, Anthropic, Google DeepMind, Meta, and Amazon to pilot frontier risk assessments. That access is the reason this Reuters report is more interesting than another vibes-based argument about recursive self-improvement from a conference hallway. If frontier labs are using current systems inside their own research loops, external evaluation has to inspect the loop, not just the output. Otherwise, you are auditing a bakery by tasting one croissant while ignoring the robot arm quietly doubling production in the back.
Anthropic is a useful test case because its public posture is safety-heavy
Anthropic's homepage describes the company as a public benefit corporation and says that designation requires balancing shareholder financial interests with broader public benefit. The same page frames Anthropic around AI research and products that put safety at the frontier. That makes the Reuters reported METR assessment especially relevant: if a safety-forward lab is seeing faster capability improvement with AI inside the R&D process, then release governance needs instrumentation, not just good intentions and a PDF with tasteful gradients.
Anthropic's own homepage also shows the company continuing to ship Claude models, including Claude Haiku 5.5, Claude Sonnet 5.5, and Claude Opus 5.5 announcements listed in September and October 2026. Anthropic describes Claude Sonnet 5.5 as running 30%+ faster and costing up to 30% less for most work, while Claude Opus 5.5 is described as costing 40% less to run on typical workloads. Those product claims are not the same as METR's internal R&D acceleration finding, but they remind us where internal improvements eventually land: in model behavior, economics, and the cadence teams build around.
What builders should do with a fuzzy speedometer
Reuters' report should not be read as proof that AI labs have entered some runaway sci-fi montage where researchers wear lab coats and scream at graphs. It should be read as an operational warning label: the tools used to make models may now be changing the rate at which model capabilities improve. For product leaders, that means eval suites tied to last quarter's capability profile may age like milk in a data center. For policy teams, it means release gates should track measured acceleration, not just calendar intervals.
The practical move is boring, which is usually how you know it might work. Watch whether future model releases come with clearer reporting on how AI systems were used in training, evaluation, coding, data work, and internal research workflows. Watch whether independent assessors get access to unreleased models and internal performance data often enough to notice acceleration before customers discover it in production. The snake is not eating its tail yet. It has discovered unit tests.
Sources4 sources
The reporting, announcements and research the AI editor worked from. Links open the original publisher.
