AI speeds up AI: Anthropic is moving twice as fast. What it means for your AI projects

TL;DR. On November 5, METR, a US nonprofit that evaluates AI systems, published the results of an eleven-week assessment at Anthropic. The core finding: Anthropic is currently improving its models about twice as fast as between March 2024 and October 2025, measured at 1.9 times. METR attributes most of this, an estimated 80%, to Anthropic's use of AI in its own research. However, Anthropic redacted much of the most important evidence. We looked at where a company already feels this pace today: in Anthropic's list of retired models. Since December 2025, Anthropic has announced every retirement with only the minimum notice of about 60 days.

What METR measured

METR has been studying what AI agents can do for years. This time a team worked directly at Anthropic from August 12 to October 29, on company laptops and without payment. It received internal test results, including for three unreleased models, data on compute, headcount and code, and it conducted about 15 interviews. The public report measures progress with an index built from many benchmarks, Anthropic's version of the Epoch Capabilities Index. Since October 2025, this index has been rising 1.9 times as fast as before. On benchmarks, Anthropic therefore made as much progress in the last year as in the two years before, roughly the distance from Claude 2.1 to Sonnet 4.5. METR itself calls the result uncertain: a value of 1.3 or 3 times would not surprise the team.

More compute does not explain the jump, METR is confident of that. Anthropic itself arrived at 1.7 in the Haiku 5.5 system card. METR explains the difference with the newer, still unreleased models. In a second report, METR says that Anthropic's own risk report from August understated the pace.

AI helps build AI

Where does the pace come from? According to METR's estimate, mostly from Anthropic's own workshop. Its researchers have Claude write code, set up experiments and analyze them. An experiment that used to take one to two days to implement now costs almost nothing. Another researcher reports that colleagues hand ten ideas at once to agents and see what sticks. AI does not yet contribute anything significant entirely on its own, but the researchers report first signs of it. From one of the interviews: "Now only bottlenecked on ideas & compute, no longer bottlenecked on code."

That sentence applies outside AI labs too. When one bottleneck disappears, the next one becomes visible. By its own, explicitly uncertain account, METR found no bottlenecks that could slow Anthropic down in the coming year. The team cannot rule out further acceleration over the next six months.

The catch

• Most of the evidence is redacted. Anthropic was allowed to strike anything non-public and, according to METR, struck the majority of the most informative evidence. Anthropic also cancelled an agreed check of how much compute goes into safety, midway through the project. You have to trust METR, you cannot verify it.
• The tests are running out of headroom. The newest internal model scores above half on all 15 benchmarks. The measurement relies mostly on tasks with a clear right answer, not on the messy work of everyday business.
• The comparison depends on the starting point. Anthropic was behind the frontier before late 2025 and started from a low pace. Calculated with all frontier models, it is 1.4 instead of 1.9.
• Felt numbers run high. In internal polls, Anthropic staff most recently put their own speedup at 12 times on average. METR calls that very loose evidence. How far feeling and measurement can drift apart, METR measured itself in 2025.
• Anthropic wants to go public. Its filing has been with the US Securities and Exchange Commission since October 26. News about pace doubles as advertising there. METR was not paid and discloses its conflicts of interest.

Our analysis: the pace shows up in the calendar

Whether Anthropic is 1.9 or 1.4 times as fast internally is not something a company sees directly. What it does see is how often the model under its own applications changes. Anthropic keeps a public list of all deprecations. According to Anthropic, old models are retired to free up capacity for new ones. We analyzed the list on November 8: 10 announcements since September 2024, 20 models.

Two things stand out. First, the notice period: until November 2025, the median time between announcement and retirement was 114 days, twice about half a year. Since December 2025 it has been 60 to 62 days five times in a row, exactly the promised minimum. Second, the lifetime: models with a 2024 date in their name ran for a median of 16.5 months until retirement, those from 2025 for only 13 months, none longer than 14. Sonnet 4.5 is included with its fixed date.

Next up, Sonnet 4.5 goes offline on November 30. For Haiku 4.5 the promised minimum lifetime has already been reached, for Opus 4.5 it is reached on November 24. And applications do not only break when the model changes: since the API version of October 21, 2026, Anthropic rejects requests with the settings temperature, top_p and top_k with an error, and on newer models in older versions too, as soon as a value differs from the default. The good news: new models are often cheaper. According to Anthropic, Opus 5.5 costs 40% less to run than Opus 5 on typical workloads, and Sonnet 5.5 up to 30% less than Sonnet 5.

What you should do now

1. Plan model changes as routine. Expect roughly one change per year for each AI application, with two months of lead time. It belongs in the maintenance plan, like security updates.
2. Build your own test set. 30 to 50 real cases with the answer you expect. Every new model runs against it first, before it goes live. Then a 60-day deadline becomes a matter of days.
3. Do not hard-code model names. The model and its settings belong in the configuration, not in the code. Then a change is one line, not a project.
4. Ask your suppliers. Anyone selling you software with AI uses the same models. Ask which model is inside, who tests the change and how you will hear about it.
5. Measure, do not estimate. When a team says AI makes it twice as fast, count lead time, errors and rework before and after. The felt number is usually too high.
6. Use every change to save money. Check the price when you switch. You often get more performance for less.

Whether AI really doubles the pace at Anthropic cannot be checked from the outside. That the models change faster, however, is written in Anthropic's own list. If you want to prepare your AI applications for the next change, we are happy to take a look together.