Eighty percent of the code Anthropic merges into its production codebase was written by Claude as of May 2026. That number was in the low single digits before February 2025.
On June 4, Anthropic co-founder Jack Clark and Anthropic Institute lead Marina Favaro published a 5,000-word blog post arguing that the AI industry needs a mechanism to slow down or temporarily stop development of frontier models. The reason: AI systems are approaching the point where they can improve themselves without human involvement, a threshold researchers call recursive self-improvement.
“The AI industry right now has a gas pedal, but it doesn’t have a brake pedal in the car,” Clark told CNN. “And we want to do some of the work to build that pedal.”
What makes this warning unusual is the source. Anthropic is not an outside critic. It is the company that filed for its IPO at a $965 billion valuation, the company that built Claude Mythos, the company that passed OpenAI in revenue. When a company spending billions on frontier model training tells the industry it might need to stop, the data behind the warning is worth reading carefully.
The Productivity Numbers Anthropic Declassified
The blog post is not a policy paper or a philosophical argument. It reads more like an internal metrics report that was deliberately made public. The specifics are what make it significant.
Anthropic engineers now ship 8x as much code per quarter as they did during the 2021 to 2025 period. Task completion time horizons have been doubling every four months, an acceleration from the previous seven-month doubling rate. Claude Opus 3, released in March 2024, could handle tasks taking about four minutes. By March 2025, Claude Opus 4.6 managed 90-minute tasks. One year later, that same model generation handles 12-hour tasks. Anthropic projects that Claude will complete skilled-person-weeks of work within 2026 and person-weeks of effort by 2027.
Code quality tracked a similar curve. Claude’s output was “somewhat worse than human-written code at Anthropic” in late 2025. By May 2026, the company says it reached parity. A March 2026 internal survey of 130 Anthropic researchers found that the median respondent estimated producing roughly 4x as much output with Mythos Preview compared to working without it.
One data point captures the scale of change better than anything else in the post: in April 2026, Claude shipped over 800 fixes to an internal system in a single task, reducing API errors by a factor of one thousand. Anthropic estimated that work would have taken a human team approximately four years.
Where Research Starts to Automate Research
The coding statistics are striking, but the research capabilities prompted the warning.
On code optimization benchmarks, Claude Mythos Preview achieved a roughly 52x speedup in April 2026. The same test produced a 3x speedup with Claude Opus 4 just eleven months earlier. The human baseline for that task: four to eight hours of work for a 4x improvement. Mythos surpassed human performance by more than an order of magnitude, and it did so autonomously.
Research judgment tests tell a similar story. When Anthropic gave Claude choices between research directions (picking moments where the human researcher’s decision had room for improvement), Opus 4.5 matched the human choice 51% of the time. Mythos Preview beat the human choice 64% of the time. In the first open-ended research project where Claude agents ran autonomously in April 2026, the models recovered 97% of the performance gap versus human researchers, who recovered 23% over the same one-week period. The compute cost: approximately $18,000.
The blog post acknowledges caveats. Humans still chose the research problem. Humans still built the scoring rubric. Not all results transferred cleanly to production-scale models. But the trajectory is the point, and the trajectory is steep.
GitHub’s own numbers provide external validation. The platform tracked 1 billion commits in all of 2025. By mid-2026, the weekly pace hit 275 million, a run rate of 14 billion annually. GitHub’s COO said the platform is “pushing incredibly hard” on capacity.
Three Futures, One Without a Precedent
Clark and Favaro laid out three possible futures for the industry.
In the first scenario, the capability curve stalls. Returns diminish. The binding constraint shifts from algorithms to physical supply chains: energy, compute, chip fabrication, interconnect bandwidth. Anthropic’s own assessment: “We don’t believe it’s likely.”
In the second, AI systems continue to compound efficiency gains, but humans retain direction-setting authority. This is the “100-person companies doing the work of 10,000-person organizations” future. The bottleneck shifts per Amdahl’s law: human review and organizational decision-making become the rate limiters, not engineering capacity. Anthropic flags a secondary risk here too. The sheer volume of new ideas, tools, and simulations generated by AI systems could exceed any organization’s capacity to evaluate them.
The third scenario is full recursive self-improvement. AI systems design, train, and refine their own successors. Humans shift to oversight, validation, and verification of what the blog post describes as an expanding “virtual lab” run by AI. The post is candid about the uncertainty: “We do not have good intuitions for what this world would look like.” It flags the possibility that rare misalignment in today’s models could compound with each generation of self-built successors, growing more frequent and less understood until control is lost entirely.
What Anthropic Proposed (and What OpenAI Rejected)
Anthropic’s ask is not for a unilateral pause. The blog post explicitly states that a unilateral stop is “achievable immediately, but accomplishes much less.” The proposal is a coordinated, verifiable slowdown across multiple well-resourced labs in multiple countries, triggered by specific conditions.
The difficulty is enforcement. Clark draws a comparison to Cold War nuclear arms control, noting that those verification regimes took decades to build. AI training runs, unlike missile silos, are easy to conceal. “The incentive to defect quietly is enormous,” the blog post states.
Anthropic committed to organizing conversations with policymakers, researchers, civil society, and competitor labs in the coming months. The company stated a conditional position: “We expect that we would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner.”
OpenAI responded on the same day with a direct counter: “Democratic governments, not private companies acting alone, must ultimately determine the rules, safeguards, and accountability mechanisms.” Translation: this is a problem for regulators, not an industry cartel.
The split illuminates a deeper strategic tension. Anthropic, which has positioned itself around safety since its founding, gains credibility by calling for a pause. OpenAI, which has positioned itself as the platform for the next computing era and just filed its own confidential IPO, has every incentive to frame regulation as a government responsibility rather than a shared industry obligation.
What the Data Means for Everyone Else
From an enterprise perspective, the takeaway is not that AI development will pause. Neither Anthropic nor anyone else has the leverage to coordinate that outcome in the near term. The verification problem alone makes it a multi-year infrastructure challenge, as Clark acknowledged.
The takeaway is what the data reveals about the pace of capability gain. If 80% of a frontier lab’s production code is model-authored today, and that percentage was single digits eighteen months ago, the same curve applies to every software organization within a few years. The companies that treat AI agents as a productivity experiment will soon find themselves competing against organizations where agents are the default workforce.
The blog post’s employee quotes make this tangible. One Anthropic engineer reported not writing any code personally for five months. Another noted that Claude “creates zero debt” in the small-favor economy that binds engineering teams together, but observed that “each of these is a lost bid for human collaboration.” A third captured the emotional whiplash: “On days where everything works well, I can’t help but think nothing I do matters. But then there are days where everything breaks and I realize I have no idea what I’ve been up to anymore.”
Those quotes read less like a corporate blog and more like field notes from the future of software engineering. The brake pedal debate is important. The productivity data is more important, because it arrives whether or not anyone agrees on how to slow down.