原文 · 未翻译
Anthropic says Claude now writes over 90% of its code and wants the world to have an AI pause button
Anthropic is sharing internal data for the first time that shows how much Claude is allegedly speeding up its own AI development. At the same time, the company is pushing for the option of a verifiable, global development pause.
Anthropic has published an extensive report through its in-house Anthropic Institute, backed by previously unreleased internal data, showing how far AI systems have already come in contributing to their own development.
The core message is that recursive self-improvement—an AI system that fully autonomously designs its own successor—hasn't been achieved yet, but it "could come sooner than most institutions are prepared for."
Code output is surging, but the real productivity gain is unclear
According to Anthropic, engineers in Q2 2026 are shipping an average of eight times as much code per day as they did in 2024. More than 80 percent of the code going into the production codebase comes from Claude. Before Claude Code launched in February 2025, that share was still in the low single digits.
Leadership estimates the total share—including scripts and experimental code—at more than 90 percent. One employee is quoted saying, "it's now been ~5 months since I last wrote any code myself."
Anthropic admits that lines of code are an imperfect metric. The eightfold increase "is almost certainly an overstatement of the true productivity gain." In an internal survey from March 2026 with 130 employees, the median estimate pegged the output boost from Mythos Preview at 4x. Anthropic itself thinks the real number is a bit lower and points to recent METR research showing that developers tend to overestimate AI productivity gains.
On code quality, the report states, "Claude-written code was somewhat worse than human-written code at Anthropic in late 2025, is roughly at parity today, and we expect it to be strictly better within the year."
An automated Claude reviewer would have caught about a third of the bugs behind past incidents on claude.ai before they hit production, according to a retrospective analysis. In another example, Claude delivered more than 800 fixes in April 2026 that cut a class of API errors by a factor of 1,000. A human would have needed four years for that work, according to the engineer in charge.
According to Anthropic, the duration of tasks that AI systems can reliably handle on their own is now doubling roughly every four months, down from seven. In March 2024, Claude Opus 3 could manage tasks in the four-minute range. A year later, Claude Sonnet 3.7 handled an hour and a half. Claude Opus 4.6 now tackles 12-hour tasks.
METR found that Claude Mythos Preview could work for "at least" 16 hours and was "at the upper end of what [METR] can measure without new tasks." If the trend holds, day-long tasks could be within reach this year, with week-long tasks following in 2027, Anthropic says.
Claude is closing in on human-level judgment in research
Beyond raw code output, Anthropic is showing progress in AI-assisted research. In an internal optimization test where Claude has to make training code as fast as possible, Claude Opus 4 hit an average speedup of about 3x in May 2025. A year later, Mythos Preview reached roughly 52x. An experienced human researcher would need four to eight hours to reach 4x.
In an analysis of real research sessions at Anthropic, the company looked at 129 moments where human developers took a suboptimal detour. Claude Mythos Preview suggested the better next step in 64 percent of those cases, up from 51 percent for Claude Opus 4.5 six months earlier. Anthropic calls this "an early signal that AI systems are getting better at making the kinds of judgment calls that AI research depends on."
The last gap: "research taste" and the architecture question
The critical bottleneck, according to Anthropic, is what the company calls "research taste:" the ability to pick the right problems and spot dead ends early. "The comparative advantage of humans as of right now is still in seeing the bigger picture and thinking beyond the confines of the immediate task," one employee is quoted as saying.
Whether that leap is even possible with today's methods remains an open question. "It is genuinely unclear whether today's training methods and architectures could unlock that capacity," the report states.
The company puts this in perspective, though: paradigm shifts like the Transformer architecture come years apart. Most progress in between is incremental work, exactly the kind of workflow Claude now handles well. Riffing on Edison's famous line about genius being one percent inspiration and 99 percent perspiration, Anthropic writes, "We see perspiration becoming increasingly automated."
Even if Claude never develops good research taste, a conservative reading of the data implies a compounding acceleration: every engineer contributes far more work than before because humans only handle the single-digit percentage of directional decisions.
Three scenarios, two of which worry Anthropic
Anthropic sketches out three future scenarios. In the first, the trend stalls.Maybe the exponential curves turn out to be S-curves, or energy and chip bottlenecks slow things down. Anthropic considers this unlikely since no flattening is visible so far.
In the second, efficiency gains continue but humans keep directional control. Companies of 100 people could do the work of 10,000 or 100,000. Anthropic sees itself on this path but warns of risks like authoritarian surveillance and tailored manipulation campaigns. Amdahl's Law kicks in, too: at Anthropic, human code review has already become the new bottleneck.
The third scenario describes full recursive self-improvement, where the pace of progress is limited only by compute. Whether the alignment problem can be solved in that case is "something we are least certain about." Rare cases of misalignment could compound, "growing more frequent but less understood until we lose control of them."
Anthropic makes the case for a verifiable development pause
The report's most striking passage concerns Anthropic's stance on slowing AI development. "We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology," the report states.
Anthropic says it would slow down or pause itself if other developers at or near the frontier did the same in a verifiable way. The Anthropic Institute plans to research and build systems that would make a credible pause possible - mechanisms that let frontier developers verify that others have actually stopped or slowed down.
The hurdles are enormous, though. Training runs are far easier to hide than missile silos. Their inputs are general-purpose, and the incentive to keep going in secret is massive: "Whoever continues while others pause could inherit the lead." A comparison to the INF Treaty for nuclear weapons seems obvious, but those verification regimes took decades to build. "We don't have that long," Anthropic writes.
A unilateral pause by a single lab would be easy to pull off right away but would accomplish far less. It would only change who's in the lead without creating the broader deliberation process that's missing. In the coming months, Anthropic plans to organize talks involving policymakers, researchers, civil society, and other AI companies.
The debate around an AI pause already came up years ago, but those calls didn't gain traction. In hindsight, the push seemed premature given what those systems could actually do. Whether this is fear-based marketing, as critics already accused Anthropic of with Mythos, will probably only become clear in retrospect.
AI News Without the Hype – Curated by Humans