An independent audit has found that Anthropic, the company behind the Claude chatbot, is now improving its artificial intelligence models about twice as fast as it was a year ago — and most of that extra speed comes from AI itself helping to build better AI.
The assessment was carried out by METR, an independent group that evaluates AI systems. Its researchers spent eleven weeks inside Anthropic, from 12 August to 29 October 2026, studying how the company’s research teams work and how quickly its models are improving. The report was published on 5 November 2026.
The headline finding is striking. Since October 2025, Anthropic’s progress on benchmark tests has run at about 1.9 times its earlier pace. In plain terms, the company is now packing roughly two years of its old rate of progress into a single year.
What the auditors actually measured
METR did not simply ask Anthropic how fast it was moving. The team worked with internal benchmark scores, training data, process records, interviews with staff and forecasts, and it measured progress using Anthropic’s own version of the Epoch Capabilities Index, known as AECI.
Before October 2025, Anthropic’s score on that index was rising by about 15 points a year. After that point, the rate jumped to about 29 points a year. Dividing one rate by the other gives the 1.9 times figure. METR also gave a range: it is 95 per cent confident the true figure lies between 1.5 and 2.8 times, and it would not be surprised by anything from 1.3 to 3 times.
The assessment looked at three internal models that the public has never seen, called Model 2, Model 3 and Model 4, whose training finished between July and October 2026, alongside public models such as Mythos Preview for comparison. METR did not get direct access to test the unreleased models itself.
AI is doing the research work
The most important question METR set out to answer was whether AI is speeding up the work used to develop better AI systems. Its answer is yes — with a careful qualification.
METR estimates that about 80 per cent of the current acceleration comes from Anthropic’s own AI assisting its researchers. Staff told the auditors that AI tools now help with coding, running experiments, analysing results and writing up findings, work that used to take human researchers far longer.
Crucially, METR is confident the speedup was not caused by simply throwing more computing power at the problem. Training compute did grow, but the auditors found convincing evidence that faster growth in computing power did not explain the jump in the rate of progress.
No, AI is not building itself yet
The viral version of this story — that AI has started improving itself without human help — is not what METR found, and the auditors say so explicitly.
AI is not yet making fully autonomous contributions to the most important parts of model training. Researchers describe only early “signs of life” for fully self-directed research, and those contributions are not yet large. The current acceleration is mainly AI-assisted, with humans still directing the work.
That distinction matters. A world in which AI helps researchers work twice as fast is already here. A world in which AI improves itself in a closed loop, with no human in the process, is not — at least not at Anthropic, and not yet.
What could slow it down — or speed it up
METR found no clear bottleneck that would stop the acceleration from continuing over the next twelve months. It looked for limits in algorithms, training compute, experiment compute and data, and found none that would obviously bind in the near term.
At the same time, the auditors are cautious about their own conclusion. Their view of the bottlenecks changed several times as they received more detailed information, and they are not confident in their current assessment. The acceleration may also be limited to well-specified tasks — the kind of work that can be measured cleanly on benchmarks — rather than every kind of research.
The finding lines up with Anthropic’s own public filings. An amended filing with the US Securities and Exchange Commission, published days before the METR report, said the company’s internal estimates indicate future models may soon meet its definition of dramatically accelerated progress attributable to AI.
The report the public cannot fully read
There is an important caveat hanging over the whole assessment. Anthropic had the right to redact evidence from the public version of the report, and it used that right heavily.
METR says the company removed a majority of the evidence it found most informative: the breakdown of progress by stage, the split between compute, data and algorithms, the bottleneck analysis, forecasts for future models and the names of specific benchmarks. A planned check of how computing resources were allocated was cancelled midway through the review.
The full report runs to about 100 pages, but it has not been published. What the public sees is a heavily redacted summary, and METR itself says the published version does not meet its own transparency standard. Readers must therefore rely on METR’s judgement more than they usually would.
Even with those limits, the audit is the most detailed independent look yet inside a frontier AI lab. It shows a company whose research engine is running at roughly double speed, powered largely by its own products — and it leaves open the question of how long that acceleration can keep compounding before it hits a wall nobody has yet identified.
Related Posts
TechnologyMicrosoft Walks Away From Oracle AI Cloud Deal as Old Nvidia Chips Fall Short — Oracle Shares Slide 4%
TechnologyTesla Cybertruck Production Ends as Cybercab Takes Over
TechnologyBYD Yangwang U9 Xtreme Becomes First Production Car to Pass 500 km/h at 503.17 km/h
TechnologyBalochistan Launches Digital Nikah Registration via NADRA and Pak ID
TechnologyNational CERT Warns Banks of AI-Powered Cyberattacks in Pakistan
TechnologyBinance Reopens Withdrawals for More Assets After $6.3 Billion Hack — HYPE, INJ, APT, NEAR and LTC Targeted by Next WeekMost Popular
How-ToPTA Tax Calculator 2026: Check Mobile Tax on Any Phone417 views
Pakistan NewsPIA Boeing 787 Lease Deal: What It Means for the Airline389 views
Canada JobsTech Jobs in Canada: Roles, Salaries, and Skills Needed387 views
SportsUEFA Champions League 2026-27: Format, Teams and Schedule384 views
TechnologyPM Shehbaz Sets November Deadline for Alibaba AI and Digitalisation Projects in Pakistan364 views
ShowbizWomen’s Fashion in Pakistan 2026: Trends, Brands and Style Guide360 views
