AI News · 5 min read
METR: AI has roughly doubled Anthropic's pace of progress
METR's Nov 5 assessment found Anthropic's models improving about 1.9x faster than before October 2025, mostly due to AI. What it found and what was redacted.
What METR published
On November 5, 2026, the AI evaluation nonprofit METR published an independent assessment of AI R&D acceleration at Anthropic ↗. The question it set out to answer: is AI now speeding up the building of AI? From August 12 to October 29, METR staff had what they call "high-access" to Anthropic's internal data. That included pretraining and reinforcement-learning metrics, training-compute figures, benchmark scores for unreleased models, forecasts, and staff interviews. The Verge ↗ and the AP both covered the findings.
The headline findings, in METR's own words and numbers:
About 2x. "As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025." METR adds it "would not be surprised" if the true figure were closer to 1.3x or 3x.
1.9x on the benchmark trend. Scores on AECI, Anthropic's version of the Epoch Capabilities Index, have grown at an estimated 1.9x the pre-October 2025 rate (95% confidence interval 1.5–2.8x). By METR's reading, that is "roughly two years of historical progress over the last year."
Not just more compute. METR says it is "confident that the acceleration was not due to an acceleration in the scaling of training compute."
Mostly AI. Its subjective estimate is that a majority, "perhaps ≈80%," of the current acceleration is due to AI. Most of that is AI assisting researchers rather than working on its own, though METR sees "signs of life" for more autonomous work.
No obvious brakes. METR found no clear bottlenecks in algorithms, compute or data over the next 12 months. It flags low confidence here, because its picture "changed several times" as it got more data.
It also reports that staff Slack polls on AI speedup reached about 12x self-reported uplift in October 2026. METR calls this "only very loose evidence." It also lists unreleased models by codename, with the newest ("Model 4") finishing training in September–October 2026. According to METR, Model 4 already scores above 50% on every AECI benchmark it was tested on, a saturation problem METR says makes it "somewhat nervous" about measuring progress from here.
The redaction problem
Anthropic was allowed to redact the public version, and it did so heavily. METR wrote on X:
"We submitted a ~100-page full report on our assessment to Anthropic. For the public version, Anthropic redacted a majority of the evidence we found most informative. As a consequence, the public must rely on our judgment more than we consider ideal."
Anthropic co-founder Sam McCandlish replied the same day. He defended the cuts:
"We redacted a lot, and I accept that this makes it harder for others to check METR's work. But we determined that sharing exactly what is driving the speedup isn't the right thing to do, both because it's commercially sensitive and because it risks handing out a roadmap for others to go faster."
He also said Anthropic will publish its own analysis of its progress speed "in the coming weeks," including where it stands against its Responsible Scaling Policy.
What people are saying
The reaction was mostly alarm. Daniel Kokotajlo, who runs the AI Futures Project behind the "AI 2027" scenario, answered McCandlish directly:
"It is pretty rich of you to cite "others could go faster" as a reason to redact so much important information when you yourself do not appear to be slowing down much at all."
Former METR researcher Sydney Von Arx said the measured pace is "slightly faster than AI 2027" (@SydneyVonArx ↗). New York lawmaker Alex Bores used the report to argue for a ban on recursive self-improvement. "The METR report below suggests they are likely right," he wrote of Anthropic staff timelines, adding: "Congress must ban RSI." (@AlexBores ↗). Apollo Research CEO Marius Hobbhahn told The Verge the findings were broadly in line with what many in the field expected, perhaps a little faster.
Our take: who should care
If you're choosing AI tools, the practical takeaway is that release cycles will keep shrinking. Anthropic's current flagship, Claude Opus 5.5 ↗, only shipped on September 22, and METR says the company already has newer models in internal use that it holds for a while before release. Expect the model you standardize on today to be replaced within months. Contracts, prompts and evals should assume that.
Three things to keep in mind:
This is one lab, measured by its own index. AECI focuses on verifiable, well-specified tasks. METR itself warns the result could be "overfitting to the benchmark distribution." It's strong evidence, not proof that AI is improving itself without people.
The redactions matter. The most informative evidence is hidden, so outsiders can't check the 2x figure independently. Treat it as METR's informed judgment, which METR openly says it is.
For buyers, nothing changes this week. No new model or price came out of this. But if you're betting a workflow on one vendor, our ChatGPT vs Claude vs Gemini comparison and best AI for coding guide are the places to keep alternatives warm.
Our view: this is the most concrete public evidence yet that frontier labs are using their own models to speed up model development. That makes the coming Anthropic report, and whether OpenAI and Google allow similar audits, worth watching more closely than any single benchmark launch.
Sources
METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 ↗, November 5, 2026
The Verge: Why a new report on Anthropic automating AI development is alarming researchers ↗, November 7, 2026
METR thread on X ↗ and Sam McCandlish's response ↗, November 5, 2026