Risk signalResearch assessmentNot an incident

Anthropic and OpenAI report faster AI research driven by their own agents

Anthropic and OpenAI report that AI agents now perform a substantial share of their own research and development work. Anthropic says Claude authored more than 80 percent of the code merged into its codebase in May 2026; OpenAI describes its system as an automated research intern and is targeting an automated AI researcher by March 2028. Both companies also stress that full recursive self-improvement has not been achieved and that safe human control remains unsolved.

Published: Sep 19, 2026
Signal type: Research assessment
Confidence: 99%
Organization: Anthropic, OpenAI
Sources: 3
Last reviewed: Sep 20, 2026

Assessment

What this signal changes

The two frontier labs publish separate internal measurements pointing in the same direction: agents are performing more execution inside AI development while humans retain goal-setting, selection, and release decisions. Anthropic reports that more than 80 percent of production code merged in May 2026 was attributed to Claude, that merged code per engineer increased eightfold, and that performance on open-ended tasks is rising. OpenAI reports 3.1 agent-workdays per human workday in its research organization, an achieved automated research intern, and a March 2028 target for an automated researcher. Both companies explicitly connect this acceleration to unresolved control, monitoring, and alignment questions.

Evidence boundary

What the evidence does not show

The metrics are self-reported by the companies and have not been independently audited. Lines of code, runtime, and internally classified task success are imperfect proxies for actual research progress; Anthropic itself warns of possible overstatement, and OpenAI calls its measurements preliminary. Humans still set research goals and decide whether to scale, pause, or deploy. Neither Anthropic nor OpenAI claims to have achieved full recursive self-improvement. This signal is not an autonomous incident and does not change incident counts or the incident timeline.

Classification: This page records a sourced warning, research finding, policy action, or industry response. It is deliberately separate from the Incident register and does not prove that an unauthorized AI action occurred.

Evidence

Sources and context

Research

When AI builds itself

Anthropic Institute · Sep 18, 2026

Primary research report combining public benchmarks with previously unreported internal Anthropic data. It states its methods and limitations, but the internal metrics have not been independently audited.

Open original source

Primary source

Research acceleration: The view inside OpenAI

OpenAI · Sep 6, 2026

Primary company publication with internal usage, runtime, and task measurements and a clear distinction between an automated research intern and full recursive self-improvement.

Open original source

Secondary source

Will AI models achieve the ability to improve autonomously? Leading labs say the scenario is near

Associated Press · Sep 19, 2026

Independent news-agency synthesis of the Anthropic and OpenAI publications. It confirms the timing and separates current research automation from full self-improvement that has not yet been achieved.

Open original source