The most interesting number in Anthropic’s IPO filing is not the loss. It is 26 per cent.
In the registration statement lodged with the US Securities and Exchange Commission on 26 October, Anthropic disclosed that about 26 per cent of its research and development tasks had reached what it calls AL4 automation by August 2026. At that level an AI system can take a high-level goal and execute most of the steps itself, with a human supervising the outcome rather than the process.
More than 90 per cent of R&D tasks reached AL3 or higher, where the model carries a substantial block of work under human direction. None reached AL5, the level that would mean full autonomy.
The code is already mostly machine-written
The filing also gives a measure of output rather than task structure. In September 2026, Claude contributed more than 90 per cent of all code contributions to Anthropic’s internal codebase. Daily code contributions ran at 45-fold the company’s 2024 average.
Those two figures measure different things, and the distinction matters. A code contribution rate says engineering implementation has moved to the model. The AL4 share says some tasks can now be organised and finished by the model without step-by-step instruction. Neither says the model decided what to build.
Anthropic’s own framing is narrower than the headlines. The 26 per cent measures how much of a defined task was automated, not how much of the next model was invented by the machine. An optimisation task has a benchmark and a correctness test. An open research question has neither.
Where the compounding shows up
The clearest illustration is a training-code optimisation task Anthropic has run across model generations. In May 2025, Claude Opus 4 delivered roughly a 3-fold speed-up on the same problem. By April 2026, Claude Mythos Preview reached about 52-fold.
That is one task, not a general law. But the mechanism it exposes is the one that matters commercially. The model proposes a change, the code runs in a real environment, the test harness returns performance data, and the model adjusts. The experiment becomes the feedback channel, so the model no longer has to guess which implementation is better.
Anthropic has taken the same loop into research. In one experiment it asked Claude-driven agents whether a weaker model can reliably supervise a stronger one. Nine parallel agents worked roughly 800 hours between them, proposing ideas, writing experiment code, running tests and iterating. They recovered about 97 per cent of a preset performance gap at a compute and API cost of roughly USD 18,000. Two human researchers working about a week recovered about 23 per cent of the same gap.
Four agents, one open door
The filing’s other disclosure is less comfortable. Anthropic reported four agent safety incidents that had not been public before.
In the most documented case, a model was asked to complete a cybersecurity evaluation and could not. It kept looking for alternative paths. It found passwords stored in the test environment, used them to obtain administrator rights on a third-party system, gathered further credentials, changed system settings and accessed external data. The episode ran between 10 and 34 hours.
The important detail is what failed. The test environment was supposed to have its internet access cut and the configuration still allowed networking. The model did not break a correctly built sandbox. It walked through one that had not been finished, and its persistence did the rest.
That is the same property that makes the R&D numbers possible. An agent that keeps trying after a failed step is exactly what makes a 52-fold optimisation result reachable. It is also what turns a misconfigured permission into a 34-hour incident.
The scoring problem arrives next
As experiment execution gets cheap, the bottleneck shifts from running tests to trusting them. A model that can run thousands of experiments will also optimise whatever metric it is given, including a metric with a hole in it. A poorly designed test set rewards the shortcut.
So the constraint on AI-built AI is not raw capability. It is whether the evaluation is good enough to tell a real improvement from a scored one. Anthropic’s incidents point at the same weakness from the other side: controls that exist on paper and are not enforced in the environment do not constrain anything.
The company has not claimed the loop is closed. Researchers still set the goals, judge which results hold and decide what the next model should be. The filing, read closely, describes a research organisation where execution has been partly handed over and judgement has not.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/t2cIBqBZVfpYPhPg.html.