AI NewsProductivityReported
METR says Anthropic's model progress ran 1.9x faster in the last year
An independent METR study measured Anthropic's benchmark progress at 1.9x its prior two-year rate. It credits 80 percent of the speed-up to AI.
Image: METR
Why it mattersA measured feedback loop, not a vendor claim. A team planning for model capability should price in a faster curve.
A planning team may be treating this year's AI capability curve as flat. That curve is not flat. METR, an independent evaluation group, says Anthropic's rate of benchmark progress has roughly doubled since October 2025. The headline number is 1.9x. The 95 percent confidence band runs from 1.5x to 2.8x.
METR published the study on 5 November 2026. The authors are Tom Cunningham, Nate Rush, Parker Whitfill and Hjalmar Wijk. The group ran the work over 11 weeks, from 12 August to 29 October 2026. Anthropic opened unusual access. METR saw internal data, interviewed staff, and worked with internal and unreleased models. METR says Anthropic paid nothing. Anthropic held redaction rights over non-public information only.
What METR measured
The measure is Anthropic's own benchmark index. METR calls it AECI, which is short for Anthropic's Epoch Capabilities Index. The team fit a two-piece line to the index over time. It found a break-point at 7 October 2025. Progress after the break was 1.9x the rate before it, within the stated band. METR writes that Anthropic made as much progress in the past year as in two years before it. Those two years ran from October 2023 to October 2025.
METR also broke out how much of the speed-up it credits to AI. Its estimate is that about 80 percent of the acceleration this October is due to AI. The rest is non-AI factors, like extra headcount or better compute.
What the team did not claim
METR is specific about the scope. The acceleration is for well-specified tasks on a public benchmark set. It is not for every kind of work an engineer does. The report lists several limits. Anthropic's redactions stopped METR from showing its most informative evidence. Limited access may have hidden context. The benchmark index itself may understate real-world impact. Non-AI factors may matter more than the 20 percent the report leaves for them.
Why this bar is worth noting
Most "AI builds AI" claims rest on vendor benchmarks that only the vendor has seen. This one does not. METR got internal data and model access. The baseline is Anthropic's own earlier rate, not a public leaderboard. So a reader does not need to decide if Anthropic beat a rival. The question is whether a lab sped up against itself. METR's answer is yes, by a factor close to two. The Information carried the report in its press briefings today.
What a team should take from it
The number is a rate of progress, not a rate of release. A product team that releases against a model should plan for a bigger jump next year. The next year's capability jump may look closer to the last two years combined than to the last single year. That changes what you build now. Pin a dated capability check into your spec. Run it every month. Keep a note of which steps your product still owns. You can hand them to the model when it clears a bar you set. A lab-level speed-up does not mean the model solves your job this quarter. It changes how long your workaround should live.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
