METR AI Progress Acceleration Study: Anthropic at 2x

What the METR AI progress acceleration study actually found

Until recently, the pace of frontier AI was commonly explained by hardware: more chips, bigger training runs, steadier gains. Then, on 5 November 2026, the nonprofit METR published its AI progress acceleration study of Anthropic. It reports that the lab's models have been improving about twice as fast since October 2025 as they did from March 2024 to October 2025. Two years of progress, compressed into one.

If you run a business, this matters now and not in some distant strategy cycle. The reason for the speed-up is more uncomfortable than "they bought more GPUs", and the evidence behind it has holes you should know about before you quote it to your board. Where the number comes from, how much to trust it, and what it changes in your next budget all come further down.

The Anthropic case in numbers

METR is an independent evaluation group. Its staff worked inside Anthropic from 12 August to 29 October 2026, with access to internal figures on pretraining, reinforcement learning, compute use and benchmarks, plus time with the researchers themselves. The result is one of the few outside looks at how a frontier lab's pace is changing. It is also the closest thing to a documented case of AI helping to build the next generation of AI.

The benchmark trend: roughly 1.9x

METR tracked progress with AECI, Anthropic's internal variant of the Epoch Capabilities Index, which folds many benchmark scores into a single number. Growth since October 2025 came out at about 1.9 times the earlier slope. The 95% confidence interval runs from 1.5x to 2.8x. In METR's own framing, a 2x acceleration means two years of historical progress arrives in one.

Think of a metric that used to climb 15 points a year and now climbs 30. Nothing about the finish line changed. The runners simply began covering the track at double speed.

The attribution: about 80% from AI

METR's judgment is that roughly 80% of the current acceleration comes from Anthropic's use of AI in its own research, and that the share is growing. It labels this estimate as carrying substantial uncertainty. The more important claim is what it rules out. METR says it received convincing evidence that the speed-up was not caused by faster growth in training compute. More chips are not the story.

The self-reports: from 1.2x to 12x

Anthropic also shared survey data on how much staff feel AI speeds up their work. Average self-reported uplift rose from about 1.2x with Claude 3.7 Sonnet to about 12x for the most capable internal model in October 2026. METR fitted the trend at roughly 3.9x per year. It also handled the figure carefully, as we will see in the next section.

Two years of progress folded into one is not a forecast. It is a measurement, taken by someone outside the building.

Where the evidence is soft

A leader who forwards this headline without the caveats will eventually be corrected by someone in the room. The caveats are substantial, and METR states most of them itself.

The range is wide. The headline is about 2x, but METR says the true figure plausibly sits anywhere from 1.3x to 3x. Treat "twice as fast" as a central estimate and not a fixed constant.

The 12x figure is loose evidence. METR points out that self-reported uplift can overstate true productivity. People are poor judges of their own speed. The interviews were also mixed: answers varied, with many researchers pointing to a speedup of around 2x, while others were unclear or disagreed.

Much of the underlying evidence is hidden. According to coverage of the report, Anthropic redacted the majority of the material METR considered most informative. Outside readers see the conclusions without the data that produced them. That limits independent verification.

The yardstick is nearing its ceiling. Reporting on the assessment notes that Anthropic's latest internal model already scores above 50% on every test in the index. A version of the index that includes rival labs' models shows only about a 1.4x speedup. Whether the acceleration is an Anthropic-specific effect or a wider industry pattern is still open.

The scope is narrow. METR measured improvement on well-specified tasks. Humans remain in the loop, and most of the acceleration comes from researchers using AI as an assistant, not from AI running research alone. The report does not claim that recursive self-improvement has arrived. It does say METR found no convincing near-term bottleneck in algorithms, compute, experiment capacity or data.

One more detail deserves attention. Anthropic's own Responsible Scaling Policy treats a doubling of AI progress, substantially attributable to automated research, as a warning-level trigger. METR considers it plausible that this line was already crossed and that Anthropic's public August 2026 risk report understates the acceleration. The finding therefore matters for governance as well as for engineering. Investors watching the company's IPO preparations will read it that way, which we covered in our analysis of Anthropic's IPO valuation.

What the METR AI progress acceleration study means for your roadmap

Your company is probably not training frontier models. The implications for you are indirect, but they are practical. They come down to three shifts.

Vendor capabilities will move faster than your contracts

If a leading lab's capability gains run near 2x its earlier pace, model generations will arrive faster than the typical 12-to-36-month software contract expects. A system you approve in January may be two generations old by the following winter. That favors short commitments, usage-based pricing and architectures where the model is a replaceable part. For a structured way to compare suppliers, see our checklist for choosing an AI agent vendor.

The planning horizon shrinks

A three-year automation plan built on today's agent performance is built on sand. Plans that hold up are written in quarters, with decision points tied to measured results. The question stops being "which tool do we buy" and becomes "how quickly can we swap, test and scale". A small business has an edge here, because it can change course faster than a company with a six-month procurement cycle.

Capability is not the same as reliability

Faster model progress on benchmarks does not automatically mean faster results in your operations. Our piece on real-world AI agent performance shows how large the gap can be between a benchmark score and finished work. METR's finding speaks about the speed of the frontier. Whether the tool in your accounts payable queue behaves reliably is still a local question, and only local testing answers it.

Measure your own speedup before trusting anyone's

The most transferable lesson from the Anthropic case is methodological. METR defines speedup as the ratio between a person's or team's output with AI and without it. It also admits that output is hard to define and that surveys inflate the answer. Copy the definition and drop the surveys.

A workable approach for a mid-sized company:

• Pick one process with a countable output. Purchase-order approvals, contract reviews, invoice matching, support tickets resolved.
• Record the baseline for four weeks. Cycle time, error rate, hours of senior attention consumed.
• Run the AI-assisted version on a comparable slice. Do not rely on how fast it feels. Compare the numbers.
• Divide. If a weekly approval queue took 40 hours of manager time and now takes 10, your speedup is 4x. (This is a hypothetical illustration of the arithmetic, not a client result.)
• Repeat quarterly. METR's central lesson is that the rate itself changes, so a single measurement ages quickly.

Some measured results are public. Sophos reported a 96% cut in threat investigation time, and Basis halved its tax workbook time. Both are useful reference points. They are not your numbers, though, and the board will ask for yours.

Five moves for the next two quarters

The assessment does not require panic. It does require a sharper posture. These five moves follow from it, and they reflect my interpretation of the findings and not METR's recommendations.

1. Shorten your AI commitments. Prefer contracts of 6 to 12 months, or ones with exit clauses tied to capability benchmarks.
2. Separate the model from the workflow. Keep prompts, tools, permissions and data connections in your own layer, so changing the engine does not mean rebuilding the car.
3. Put one executive in charge of AI measurement. Not enthusiasm, not pilots-for-pilots'-sake: one owner of the speedup numbers.
4. Build agent governance now. Faster models mean more capable agents with more access. Our guide to managing AI agents in the enterprise sets out the minimum structure, and the analysis of Anthropic's agent-control limits explains why it cannot wait.
5. Rehearse the quarterly review. Once a quarter, ask what changed in the capability frontier and which of your processes are now cheaper to automate than last time.

Leaders who do this tend to describe the same shift. The feeling of chasing every headline gives way to calm and to real control: you know which numbers you track, which decisions are reversible, and which are not. The external perception changes too. A board or an investor who sees a CEO bringing measured speedups and a clear exit plan, instead of excitement about the latest model, reads that executive as someone who turns a fast-moving technology into predictable profit.

FAQ

What is the METR AI progress acceleration study? It is an independent assessment published by the evaluation nonprofit METR on 5 November 2026. After 11 weeks inside Anthropic, METR concluded that the lab's model capabilities are improving at about twice the rate of March 2024 to October 2025. It estimates that most of the speed-up comes from Anthropic using AI in its own research.

Does the study prove that AI is improving itself without humans? No. METR describes a pattern where humans use AI as a research assistant, and human researchers remain in control. The report does not claim a closed loop of autonomous self-improvement, though staff told METR they expect autonomous contributions to grow.

How reliable is the 2x figure? It is a central estimate with real uncertainty. The 95% confidence interval for the benchmark trend is 1.5x to 2.8x, and METR says the true value could plausibly range from 1.3x to 3x. Anthropic also redacted much of the underlying evidence, so outside verification is limited.

Was the acceleration caused by more computing power? According to METR, no. It says it received convincing evidence that the acceleration did not come from faster growth in training compute. It attributes about 80% of the current speed-up to AI-assisted research, with a high degree of uncertainty.

Does this mean my business tools will get twice as good in a year? Not necessarily. The finding covers progress on well-specified benchmark tasks inside one lab, not performance in your processes. Frontier capability will likely keep rising quickly, but you only learn what it does for your operations by measuring a specific workflow.

What should a business leader do first? Choose one process with a countable output, record a four-week baseline, and compare it with an AI-assisted version. That gives you a speedup figure of your own, one the board can check, and it tells you where deeper automation pays off. If you want to talk through which process to start with, you can ask our AI assistant.

The METR report settles a question that used to live in forecasts: AI is already shortening the road to better AI, at least at one frontier lab. You cannot control that pace, but you can decide how much of it reaches your own balance sheet. The companies that win this decade will not be the ones with the boldest AI announcements, but the ones that can show, quarter after quarter, a measured speedup in their own operations.