METR says Anthropic's models now improve 1.9 times as fast, and Anthropic redacted the evidence
METR's eleven-week look inside Anthropic found two years of model progress packed into one, mostly from AI doing the research itself. The strongest evidence is redacted, days before an IPO investors now price nearer $1 trillion.
Vincent Jiang · 3 min read
Three models nobody outside has seen
Between July and October, Anthropic$350B — Anthropic, private, latest valuation $350B finished training three frontier models no outsider has used, one in July, one by September, one in early October. The inventory comes from METR, a Berkeley nonprofit evaluator whose staff worked inside Anthropic's data from August 12 to October 29 12. In internal Slack polls, staff put their productivity gain from the most capable hidden model near 12x, a figure METR itself calls very loose evidence 1. That stockpile is the engine a trillion-dollar listing would sell.
Twice the pace, and not from compute
Since October 2025, scores on AECI, Anthropic's internal version of the Epoch Capabilities Index, have climbed 1.9 times as fast, with a 95% confidence interval of 1.5–2.8x, moving from roughly 15 points a year to 29: two years of progress in one 13.
METR confidently rules out faster training-compute scaling, guesses about 80% of the current acceleration comes from AI doing the R&D itself, and sees no bottleneck in algorithms, compute or data for twelve months 1. Weeks earlier, CEO Dario Amodei had urged the industry to "slow the pace at which we improve the capabilities of AI models" 4.
Six estimates of Anthropic's speedup; only the top one clears its own 2x line
Direction settled: faster. Magnitude, and whether it trips the red line, is not.
Data
| Value | |
|---|---|
| METR, easy-tail variant | 1.1x |
| METR incl. rival models | 1.4x |
| Opus 5.5 system card | 1.5x |
| Haiku 5.5 system card | 1.7x |
| METR, main estimate | 1.9x |
| METR, hard-tail variant | 2.9x |
The proof sits behind the redactions
Anthropic held unilateral redaction rights and cut "a majority of the evidence we found most informative" for METR's risk conclusions, describing the cuts as aimed at information that could accelerate other developers 1.
METR adds that Anthropic's public August risk report understates the acceleration; a companion review calls the company's case that it stayed under its own 2x threshold "overstated"; and the November 2 amended IPO filing says future models "may imminently meet" that threshold 34.
Co-founder Sam McCandlish, Anthropic's responsible scaling officer, concedes the cuts make METR's work harder to check and calls the driver of the speedup commercially sensitive 4. Buyers are underwriting a speedometer they cannot open.
Who pays this month
Anthropic filed October 26 for a Nasdaq listing under ticker ANT 5, with a roadshow planned for the week of November 9 and an offering some expect near $100 billion, down from early talk of $2 trillion 6.
After Thursday's $2.5 trillion tech rout, three major investors told The Wall Street Journal that $1.5 trillion is unlikely, one putting the number nearer $1 trillion 7. Capital Group, at $10 billion, and Fidelity, at $5 billion, lead roughly $16 billion of fund money already in 8, while Michael Burry counts $537 billion of compute commitments and calls the boom "catastrophically vulnerable" 5.
The IPO price tag has halved in three weeks of talk
Data
| Reported valuation | |
|---|---|
| Early talk | $2T |
| Investors tell WSJ | $1.5T |
| One investor's number | $1T |
The reading that decides it
The counterweight is METR's own spread: include rival labs' models and the acceleration reads just 1.4x 3. Anthropic must decide within days whether to launch, with October revenue still blank in the amended filing 67. Next on the record are the six METR staff embedded under the October 15 arrangement, with employee-like access and publication Anthropic cannot veto outright 39. Watch whether a $1 trillion price clears on conclusions alone 39.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
