Skip to content
NewsResearch

Anthropic Publishes Internal Metrics Showing Claude Now Writes Over 80 Percent of Its Production Code

· by Pondero Newsdesk

The short version

The Anthropic Institute released previously unreported internal data on September 18, 2026 showing Claude authors the majority of code merged to Anthropic's codebase, engineers ship eight times more code per day than in 2024, and the task horizon for autonomous AI agents has been doubling roughly every four months.

Anthropic Publishes Internal Metrics Showing Claude Now Writes Over 80 Percent of Its Production Code

As of May 2026, Claude authored over 80 percent of the code merged into Anthropic's production codebase. That figure comes from "When AI builds itself," a report the Anthropic Institute published on September 18, 2026, which is the first time a major AI lab has released detailed internal data on how far recursive self-improvement has actually progressed in its own engineering and research workflows.

What the data shows

The 80 percent share covers code attributed to Claude across lines merged to production, per the Anthropic Institute report. That is the conservative figure. Anthropic leadership separately estimated publicly that 90 percent or more of all code written at the company, including scripts and experimental code, comes from Claude; the difference reflects attribution gaps and auto-generated artifacts not counted in the production-merge measure. Before Claude Code launched in research preview in February 2025, the Claude-authored share was in the low single digits.

The productivity effect is measurable. Lines of code merged per engineer per day held flat through Anthropic's first four years (2021 through 2024), then began rising in 2025 when Claude started running and testing code rather than just suggesting it. The slope steepened again in 2026 when models began handling longer autonomous work sessions. By the second quarter of 2026, the typical Anthropic engineer was merging eight times as much code per day as in 2024, per the report. That jump reflects engineers directing and reviewing Claude's output rather than writing code themselves.

The report also tracks the task horizon: the length of task an AI agent can complete reliably on its own. That horizon has been doubling roughly every four months, accelerating from a prior trend of once every seven months. In March 2024, Claude Opus 3 could complete software tasks a human would spend about four minutes on. A year later, Claude Sonnet 3.7 handled tasks requiring roughly 90 minutes. Claude Opus 4.6 reached 12-hour tasks. METR, which runs benchmarks for long-duration agent performance, found that Claude Mythos Preview could sustain work for "at least" 16 hours, at "the upper end of what METR can measure without new tasks," per the report.

On task completion quality, Claude Code session success rates for all four difficulty tiers (trivial, routine, substantial, and open-ended) have converged between 88 and 92 percent as of September 2026, per a same-day update to the report. Open-ended problems showed the largest improvement: from about 26 percent to 91 percent, with the sharpest jump occurring in March 2026 after the Mythos Preview internal rollout.

For research judgment, the Anthropic team examined 129 real Claude Code sessions from January through March 2026 where researchers hit a wrong turn in open-ended investigations. Different Claude models were asked what to do at those moments, and a separate Claude judge evaluated whether the AI or the human suggested the better next step. Claude Mythos Preview chose the better option 64 percent of the time. Claude Opus 4.5 in November 2025 reached 51 percent. Both figures used a test set where human choices were known to have room for improvement, so neither is a like-for-like comparison against human judgment in general.

A separate experiment in April 2026 assigned Claude-powered agents an open AI safety problem: whether a weaker model can reliably supervise a stronger one. The agents proposed hypotheses, ran tests, shared findings with parallel agents, and iterated over about 800 cumulative compute-hours. Two human researchers working for roughly a week recovered 23 percent of the available performance gap; the agents recovered 97 percent at a compute cost of approximately $18,000, per the report. Anthropic noted the result did not transfer cleanly to production-scale models and that humans still chose the problem and wrote the scoring rubric.

Why it matters

For engineering teams calibrating when to build multi-agent pipelines or schedule autonomous overnight runs, the task-horizon curve is the most actionable number. If the doubling-every-four-months pace holds, tasks that take a skilled engineer several days could fall within autonomous AI range before the end of 2026. Tasks requiring several weeks could follow in 2027. The report treats those projections as possible rather than certain: the trend could plateau, compute constraints could slow progress, or architectures could reach a ceiling before unlocking full direction-setting autonomy.

The productivity data is also a signal about how fast AI adoption is reshaping engineering workflows more broadly. GitHub saw roughly one billion code commits across all of 2025. By mid-2026 that pace had reached about 275 million per week, on pace for roughly 14 billion across the year, per the Anthropic report citing GitHub's COO. Anthropic's internal numbers are part of that industry-wide surge.

The report is explicit about the gap that remains. Code quality from Claude is described as roughly at parity with senior human engineers today and expected to be "strictly better within the year," per Anthropic. The capability that persists as a human domain is research taste: deciding which problems matter, which results to trust, and when to abandon an approach. The early "better next step" results for Mythos Preview are the first evidence this gap is narrowing, but the measurements are narrow in scope, and the report treats them as early signal rather than a trend.

Context

The Anthropic Institute is a unit inside Anthropic that publishes research on AI's trajectory and societal implications. "When AI builds itself" was co-authored by Marina Favaro and Jack Clark, with editorial support from Santi Ruiz. It draws on public benchmarks including SWE-bench, CORE-Bench, and METR's time-horizon evals alongside internal data that Anthropic describes as previously unreported.

Anthropic stated outright that full recursive self-improvement has not been reached. The report defines full recursive self-improvement as the point where AI systems can autonomously design their own successors without human direction. The company also described its position on a potential slowdown or pause: it would support one if other frontier labs did so in a verifiable manner and a coordination mechanism existed to prevent any actor from advancing in secret during a pause. The Anthropic Institute said it would organize conversations with policymakers, researchers, civil society, and other AI companies on these questions in the coming months.

What to watch next

Three measurements will indicate how the current trajectory develops. First, whether the task-horizon doubling rate continues: METR's current eval suite is already at its upper bound, which means new task benchmarks need to be published before the trend can be tracked clearly past the 16-hour mark. Second, whether code quality crosses from parity to strictly better at Anthropic and when other frontier labs, specifically OpenAI and Google DeepMind, release comparable internal data; the current report is Anthropic-specific and covers one organization's conditions. Third, whether research judgment benchmarks continue to improve on problem selection, not just experiment execution: the 64 percent "better next step" figure for Mythos Preview is a narrow measure taken from moments where human choices had known room for improvement, and follow-up measurements will show whether the improvement generalizes.

Sources