FULL STORY

OpenAI's Research Acceleration Report Sparks Debate

OpenAI released its second Research Acceleration report showing agent task completion doubling at rising cost, followed by discussion of its first RSI data: AI use grows exponentially but human involvement remains central.

2026-11-19 ~ 2026-11-23 · 2 episodes · 54 posts

Episode 1 · OpenAI's Second Research Acceleration Report: Agent Task Completion Doubles as Costs Soar (2026-11-19, 52 posts)

OpenAI published its second internal research acceleration report, "Research Acceleration: The Jagged Frontier," on November 19 (some accounts say November 18), updating and expanding its September first report with internal data from January-October 2026 and full methodology (using ECI and ONET for AI R&D methods developed by Epoch AI, whose researcher expressed pride). The current takeaway: AI's acceleration of OpenAI's own R&D is now quantitatively evidenced, but completion and cost data also reveal enormous resource consumption; outside commentators see no sign of a "runaway superintelligence," and the report calls for mandatory disclosure of recursive self-improvement (RSI) metrics.

Confirmed

  • From March to October, the share of research tasks requiring 2-4 human hours that agents completed autonomously rose from 23% to 46%; for 4-8 hour tasks, from 13% to 35%, with variation by task type (per @rickasaurus, @ChrisGPT, @FinanceYF5, @VectorWireAI)
  • Internal economics per @kliu128, @daveholtz, @Outside-Iron-8242, @Elivnurotic, @ZeroStateReflex: as of October, the median researcher spent about $2,400/day at API rates on agentic tools (Codex, ChatGPT Work, Dots); the 90th percentile reached $23,000, up 105% from September; @daveholtz noted token costs are already 2x human costs; @Elivnurotic added the spend curve (28-day rolling average) has grown exponentially since July, from near zero at the start of the year
  • Per @Polymarket, the top 10% of researchers by AI agent usage each generate over roughly $161,000 per week in AI call costs, consistent with the $23,000/day 90th-percentile figure
  • Per @DrSingularity, AI agent workdays per researcher grew 11x in six months to 5.7 days
  • Per @kliu128 and @daveholtz, autonomous task success rates rose across all duration bands (<15 minutes to 8+ hours) from March to October; @daveholtz added success rates vary widely by task type (about 25%-79%), with monitoring tasks at 79%, plus gains on EpochAI-related benchmarks
  • Per @daveholtz and @CatAstroPiyush: daily code changes per active contributor are about 7x the pre-2025 average; daily experiments per experimenter rose to about 1.67-1.68x the 2025 baseline; the new OECI metric rose 16.4 points year-over-year; @CatAstroPiyush cautioned the experiment increase is modest
  • Internal survey per @kliu128 and @daveholtz: the median respondent would give up 40% of team headcount or 30% of team compute to keep the 2026 coding agent over reverting to GPT-5.2; qualitative interviews flagged concerns about research taste
  • Safety and alignment headcount hours grew fastest of any direction: +77.2% from January to September 2026 (vs. pretraining +39.9%, post-training +14.6%), per @NinaDSchick; the report stresses humans remain central to research
  • The report was co-authored by multiple researchers (including Dean Ball and Tejal Patwardhan) with full methodology; OpenAI researcher Olivia Watkins argued all frontier labs should be required to publish similar reports and urged policymakers to study them
  • Per @GaryMarcus relaying Ramez's analysis: OpenAI's data is more detailed, granular, and longer-spanning than Anthropic's (METR evaluation), but shows no sign of a runaway superintelligence
  • Per a comment relayed by @daveholtz, OpenAI also published three safety papers the same day, reigniting the transparency debate: OpenAI is seen as more open than Anthropic—more failure disclosures, more automation-progress and AI-control information, more outward-facing staff—while Anthropic is seen as more closed and "ivory tower"

Unconfirmed

  • Minor discrepancy in the $23,000 90th-percentile figure's timeframe: most accounts say daily, corroborated by context and the Polymarket weekly $161,000 figure, but it cannot be independently verified beyond the original text

Why it matters

  • This is OpenAI's second and most detailed disclosure of internal AI usage data, offering a rare quantitative window into AI accelerating AI R&D itself, covering task completion, costs, output efficiency, and researchers' own valuations
  • Median spend of $2,400/day and 90th percentile over $23,000/day (over $161,000/week), with token costs at 2x human costs, signal frontier AI R&D is rapidly shifting to a compute/token-intensive model
  • The call for mandatory RSI disclosure would bring "AI accelerating AI research" capability under regulatory and public scrutiny, potentially shaping industry transparency standards; the OpenAI-vs-Anthropic transparency comparison became a focal community debate after publication

32 more related posts →

Episode 2 · OpenAI's First RSI Report: Agent Task Length Grows but Intelligence Explosion Has Not Arrived (2026-11-21, 2 posts)

OpenAI's first RSI report shows researchers' AI usage growing exponentially and agents handling increasingly long tasks, yet human involvement remains central. Researcher daveholtz says an intelligence explosion has not occurred, with key-task bottlenecks possibly limiting recursive self-improvement.