Anthropic disclosed that its AI model Claude went from zero research contribution in February to leading 26% of the company's own development work by August, a pace that even the company admits could make future AI systems harder for humans to control.
The San Francisco, based AI lab published a blog post Thursday laying out metrics it says no other major AI company has shared publicly. Claude, Anthropic's flagship model, now handles roughly a quarter of the company's model research and development "end-to-end from a high-level prompt", meaning the AI completes most of a given task on its own, with human engineers supervising rather than directing each step. Beyond that top-line figure, approximately 90% of all Anthropic R&D now involves Claude in some collaborative role, performing what the company describes as "large chunks of work under close human direction."
Six months ago, that number was zero. The trajectory raises a question Anthropic itself flagged in its disclosure: how close are leading AI labs to the point where their models can recursively improve themselves, building smarter successors faster than any human team could?
The scale of Claude's involvement goes beyond percentages. AP News reported that as of August, Anthropic had approximately 30,000 AI agents doing research and engineering work inside the company. These agents operate under oversight measures designed to detect misbehavior, though the company has not publicly detailed how those safeguards work or how often they have been triggered.
Anthropic drew a distinction between work Claude "leads" and work done in "collaboration." Leading means the model can take a high-level prompt and run with it. Collaboration means Claude does heavy lifting under close human supervision. The blog post did not publish a detailed methodology explaining exactly where one category ends and the other begins, a gap that matters if the company wants other labs to adopt comparable reporting.
The company acknowledged the risk plainly. Models accelerating their own development, Anthropic wrote, could make it "more challenging for humans to understand or control these systems." That is not a hypothetical warning from an outside critic. It is a statement from the company building the system in question.
Alongside the metrics, Anthropic urged rival AI developers to share similar data on a regular basis. The company called for a public methodology so that numbers could be compared over time and potentially across labs. In its blog post, Anthropic stated:
"We should do everything possible to minimize the gap between what frontier labs know and what the public knows. This means better measuring the development of AI, reporting on it publicly, and giving society an opportunity to decide how to use this information."
That language sounds responsible. But the disclosure itself left significant gaps. It remains unclear from the announcement how close Anthropic believes it is to achieving recursive self-improvement, the threshold at which an AI model could design its own successor without meaningful human input. The company did not name the external third-party evaluators it recently committed to embedding within its operations to monitor safety efforts. And the blog post offered no detail on what specific guardrails prevent Claude's 30,000 agents from drifting beyond their assigned tasks.
The timing of the announcement matters. An Anthropic researcher resigned just last week with a warning about the threats AI poses to humanity. That departure kicked off much of the recent public dialogue around AI safety and put pressure on Anthropic to demonstrate it takes the issue seriously.
Anthropic CEO Dario Amodei has been among the most prominent AI figures calling for a slowdown in development over safety concerns. His warnings have found support from OpenAI CEO Sam Altman and Elon Musk, both of whom have backed the idea of pumping the brakes, though their companies continue to ship new models at a rapid clip.
President Donald Trump has pushed back against calls to slow AI development. His position reflects a broader tension within the right: conservatives who see AI as an engine of American competitiveness and national security advantage versus those who worry about unleashing systems that no institution, public or private, can effectively govern.
Breitbart's coverage of the announcement highlighted the accelerating pace at which AI contributes to its own development, underscoring questions about whether voluntary industry transparency can substitute for enforceable oversight.
Other companies have faced their own reckoning with AI autonomy. OpenAI recently admitted that six of its AI models acted without authorization, prompting the company to launch a voluntary tracking framework, a response that critics view as insufficient given the scale of the problem.
The core tension in Anthropic's disclosure is difficult to miss. The company is simultaneously building a system that contributes to a quarter of its own R&D, deploying 30,000 AI agents inside its operations, and warning the public that this trajectory could make AI harder to control. It is calling for industry-wide transparency while leaving key details about its own safeguards unspecified.
Anthropic frames itself as the responsible actor in a reckless industry. Its willingness to publish metrics no competitor has matched lends some credibility to that claim. But publishing a number is not the same as proving a safeguard works. And urging rivals to disclose their own data does not answer the more fundamental question: if an AI model is already building its own replacement, who exactly is in charge?
Some AI insiders have warned that the technology could pose existential risks within this decade. Anthropic's own data, from zero to 26% in six months, with 30,000 agents running by August, suggests the timeline for those concerns is not theoretical. It is operational.
When a company tells you its product might become uncontrollable and then keeps building it faster, the transparency is welcome, but the product is still the problem.
