# METR says Anthropic may have crossed its AI warning threshold in July

> An audit found faster gains in Anthropic’s model scores, with AI assistance in its research a likely factor. METR says the estimate is uncertain.

Oossa · 2026-11-06 · https://oossa.com/en/metr-says-anthropic-may-have-crossed-its-ai-warning-threshold-in-july

METR, a nonprofit that studies AI risks, says Anthropic may have reached its own warning level for AI acceleration by July 15, 2026. The organization estimates that Anthropic’s model test scores have improved about 1.9 times faster since October 2025 than they did in the period before that. Anthropic’s August 2026 public risk report said acceleration was “less than a factor of 2.” METR says the unredacted version implies a figure of 1.97, and Anthropic itself described crossing the threshold as plausible.

The score comparison uses the AECI, a combined measure of results across multiple benchmarks. Anthropic’s models gained about 15 AECI points per year before October 2025, METR says, compared with 29 points per year since. METR also examined three unreleased models that completed training between July and October 2026.

## What may be driving the faster gains?

METR estimates that around 80 percent of the acceleration comes from AI helping Anthropic’s researchers, rather than faster growth in the computing power used to train models. Employees told the auditors that some experiments that once took one or two days to set up are now “~free.” In informal October 2026 polls, staff said the latest internal model made them about 12 times more productive on average. METR calls these self-reports loose evidence that may overstate the gains; one researcher put AI’s contribution at just 15 to 20 percent.

The estimate is not a settled picture of the whole AI industry. METR says the acceleration factor for Anthropic could plausibly range from 1.3 to 3. Including OpenAI, Google and xAI in the comparison brings the estimated industry-wide rate to 1.4. METR also says AI is not yet conducting fully independent research, though it sees early signs that this could change.

## What the audit could not verify

The review ran from August 12 to October 29, 2026, with about 2.5 full-time METR staff and access to internal metrics, employee surveys and roughly 15 interviews. Anthropic could redact non-public information and removed much of the evidence METR considered most useful, including forecasts and analysis of possible bottlenecks. METR says this leaves readers relying more heavily than it would like on the auditors’ judgment.

METR says it shares Anthropic’s current assessment that acute catastrophic risk is low, but cannot rule out another sharp acceleration in the next six months. Anthropic also stopped a separate part of the project that was meant to check how the company divides computing resources between capability and safety research, saying it may revisit that audit later.

## The facts

- METR estimates Anthropic’s annual AECI score gains rose from about 15 points before October 2025 to 29 afterward.
- METR estimates AI use in Anthropic’s research accounts for about 80 percent of the acceleration.
- In informal October 2026 polls, Anthropic employees reported an average productivity gain of about 12 times with the latest internal model.
- METR says the acceleration estimate for Anthropic could range from 1.3 to 3, while its industry-wide estimate is 1.4.
- Anthropic ended the planned review of how it allocates computing resources between capability and safety research.

## Why it matters

For people following AI safety, METR’s findings suggest that AI tools may be helping researchers improve models faster, which could make a company’s warning thresholds more relevant sooner than expected. But the size of the effect is uncertain, and redactions and self-reported employee surveys limit what outsiders can confirm.

## Sources & references

1. [METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July](https://the-decoder.com/metr-finds-anthropic-may-have-crossed-its-own-ai-acceleration-warning-threshold-back-in-july/) – The Decoder, 2026-11-06

Last updated: 2026-11-06