Anthropic is improving its AI 2x faster by having its staff use AI
METR, an independent research group, spent three months inside Anthropic. It concludes the company is likely improving its AI about twice as fast as before, and that its own staff using AI is the main reason.
The short version
- METR, an independent research group, was given rare access to Anthropic's internal data from August to October.
- It found Anthropic is likely improving its AI models about twice as fast as it did before October 2025.
- Most of that speed-up seems to come from Anthropic's staff using AI in their own work, and METR saw nothing that would obviously stop it.
What METR did
METR studies how fast AI is advancing and what risks that brings. Anthropic, the maker of Claude, let METR inspect its internal records. These covered training data, test scores for unreleased models, staff surveys and interviews. METR calls the access unprecedented. Its team worked on the review from mid-August to late October 2026.
The question was simple. Is AI now speeding up the work of building better AI? If it is, progress could compound, and new abilities could arrive faster than expected.
What it found
METR tracked a combined score Anthropic uses to measure its models' skill. Since October 2025, that score has been rising 1.9x faster than its earlier trend. In other words, Anthropic made about two years of past progress in the last year.
METR is careful about the range. It would not be surprised if the true figure were anywhere from 1.3x to 3x.
Anthropic's own earlier estimate, published with one of its models, was 1.7x. METR's figure is higher because it also included newer, unreleased models.
METR saw convincing evidence that more computing power was not the cause.
METR's best guess is that about 80% of today's speed-up comes from AI. Most researchers in interviews said AI made them about twice as fast. Staff in informal polls reported much bigger gains, reaching 12x in October 2026. METR treats those polls as weak evidence.
Most of the help is AI assisting people, not working alone. Staff told METR there are signs that AI working on its own will soon play a bigger part.
The catch
Anthropic blacked out much of the evidence METR found most useful. Anthropic said this was to protect information that could help rival developers. METR is frank that the public must therefore trust its judgment more than it would like.
METR also found no clear limit that would stop further speed-ups in the next 12 months. It adds that its view on this changed several times during the review.
METR also admits it could be wrong in either direction. It may have given too much weight to factors other than AI. Or it may have missed signs that progress is speeding up even more.
What it means for companies in Indonesia
The AI tools you buy are likely to improve faster than they did a year ago. A model that falls short today may be good enough within months. One that works today may soon be overtaken. Prices and abilities can shift within a single budget year.
There is also a lesson in how the gain happened. Anthropic's speed-up came mostly from people working alongside AI every day. The same pattern is open to any company with good tools and willing teams.
What to do next
- Review your AI choices every quarter, not every year. Avoid long contracts that lock you to one model.
- Re-test tasks that AI failed six months ago. They may now work.
- Measure your own teams' gains from AI with simple before-and-after numbers.
- Ask your AI suppliers how often their models change, and how they will tell you.
- Keep people in charge of decisions as AI takes on more of the work.