OpenAI research-acceleration report
Partly confirmed
Importance: major (4 of 5)The takeaway
OpenAI said AI labs should be required to track and publicly report indicators of recursive self-improvement, while saying this does not signal an intent to pursue full RSI.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- Major (4 of 5)
- Last verified
- 19 November 2026
Your AI and this story
- GPT-6.2 Astra171 days after its cutoff
- Claude Opus 5.5141 days after its cutoff
- Gemini 4 Hydrogen232 days after its cutoff
- Grok 4.7171 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 141 days before it.
Key facts
- Task autonomy (OpenAI chart, as reported): ~46% of 2-4-hour research tasks completed without human help in Oct 2026 vs ~23% in March; success falls with task length (~35% for 4-8-hour tasks, secondary summaries)
- Jaggedness (secondary summary, unverified against the report): ~79% unassisted success on training-run monitoring tasks vs ~25% on inference and serving code
- OpenAI (X, Nov 18, 652k views): researchers’ AI use ‘continues to rise exponentially’; ‘human involvement remains central’; human effort on safety and alignment is ‘growing faster than in any other individual research area’; ‘AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI)’
- Spend (secondary reports): heavy users consume thousands of dollars of model compute per day; one write-up says research agents cost about twice researchers’ cash pay — figures not checked against the report
- Critics (DEV, commentators): figures are self-measured with no independent audit and no thresholds that would trigger action
- Same day OpenAI also published ‘Managing frontier AI risk: from training to deployment’, a recap of safeguards from early research through internal use to public release (X post 268k views), and ‘Inside the Codex sandbox’ on filesystem and network limits for coding agents
- Context: Anthropic’s S-1/A says it has met its own RSP threshold for AI-accelerated R&D (Anthropic: AI R&D threshold met (S-1/A)); the MisAnthropic Statement and Ban RSI rally targeted RSI the same week
- Mood the same day: former OpenAI researcher Will DePue posted ‘guys we’re in the fucking singularity… its all happening’ (X, 2.67M views by Nov 19), the most-viewed AI post of the day in our sweep
What happened
OpenAI released a report measuring AI’s share of its own research work. The openai.com page returned HTTP 403 to us; the numbers above come from OpenAI’s X thread and secondary write-ups that quote the report’s charts, so exact values should be checked against the original.
Why it matters
It is a frontier lab publishing a time series of how much its research is automated, the core quantity in debates about recursive self-improvement, and it explicitly asks for mandatory reporting of such indicators across labs.
Sources
6 sources from 3 sites. Numbers match the chips in the text.
6 sources: 4 primary, 1 press, 1 reaction
Primary
- OpenAI: Research acceleration: the jagged frontieropenai.com, official
- OpenAI on X: new data on how AI is changing research at OpenAIx.com, official
- OpenAI: Managing frontier AI risk: from training to deploymentopenai.com, official
- OpenAI: Inside the Codex sandboxopenai.com, official
Press
Reactions
- Will DePue on X (2.67M views)x.com, discussion
Changes
- DePue post
- Filed (primary page unreadable to us; numbers from OpenAI’s X post and secondary summaries)