FULL STORY

OpenAI's Internal AI Task Length Halves Every 86 Days

OpenAI's research automation report shows internal AI tasks now span ~15 minutes, halving every 86 days. Toby Ord and Basil Halperin's analyses sparked debate over underestimated AGI timelines.

2026-11-07 ~ 2026-11-09 · 2 episodes · 23 posts

Episode 1 · Oxford Philosopher Says Most AGI Timelines Are Too Long (2026-11-07, 3 posts)

Oxford philosopher Toby Ord noted that AI completes 80% of AI R&D tasks within a 15-minute time span in OpenAI and Google DeepMind measurements, adding that most people's AGI timelines are still too long, though he stressed the figure is only one piece of the RSI puzzle.

Episode 2 · OpenAI internal data shows 15-minute 80% time horizon with method-dependent doubling speed (2026-11-09, 20 posts)

On November 9, Basil Halperin published a thread analyzing time-horizon data for internal AI research automation tasks from OpenAI's self-reported automation report. The core finding: at an 80% success rate, the time horizon is only about 15 minutes—far below METR's headline numbers—though it is growing quickly. However, this conclusion is highly sensitive to the fitting method, and METR's baselines themselves have measurement issues.

Confirmed

  • At an 80% accuracy threshold, OpenAI's internal task time horizon was about 15 minutes as of July; using METR-style logistic fitting, time horizons at all success thresholds (25%/50%/80%) double every 58–87 days (about 2 months), much faster than the 4–7 months METR has measured publicly.
  • At the 50% threshold: naive reading gives about 10 hours with doubling every 68 days; logistic fitting gives about 6 hours with doubling every 63 days. At the 80% threshold: logistic fitting gives about 15 minutes with doubling every 86 days, while the naive reading (empirical p80) gives about 15 minutes with doubling every 585 days—a very slow rise.
  • Month-by-month logistic success-rate curves fit reasonably well (RMSE of 3.6–6.9 percentage points).
  • In OpenAI's internal data, the 80% time horizon is roughly 15x the 50% one, whereas in METR's data the 50% horizon is about 4x the 80% one—a stark discrepancy in ratios that Halperin is unsure how to interpret, speculating it may relate to low/high-context human baselines.
  • The METR paper (Kwa et al.) uses a 'low-context' human as its baseline; the paper's appendix estimates that repo maintainers familiar with the work are about an order of magnitude faster than baseline humans—switching to a high-context baseline would lower the time-horizon values by roughly 10x.
  • Kwa's (METR) cross-domain analysis shows that while absolute time horizons differ across domains (SWE-bench, LiveCodeBench, GPQA, Tesla FSD, OSWorld, MATH, etc.), growth rates are similar, doubling every 4–7 months.

Unconfirmed

  • Why OpenAI's internal time-horizon growth (doubling roughly every 2 months) is far faster than METR's public measurements. Halperin offers two explanations: (a) domain differences—internal AI R&D tasks are inherently special; (b) time-horizon progress is uneven across domains, in which case extrapolating overall AI capability trajectories from a single domain's growth rate becomes harder. He and Dylan Ryfe are working on an extended analysis.
  • Whether the 80% horizon is rising quickly depends entirely on the fitting method: the naive fit shows doubling every 585 days (slow rise), while the METR-style fit shows doubling every 86 days. Halperin explicitly emphasizes this method dependence in his summary.

Why it matters

  • Recursive self-improvement is widely discussed, but OpenAI's own data shows that at high reliability (80%), the task horizon models can complete remains very short—placing a constraint on the narrative that AI will soon autonomously carry out long-horizon R&D.
  • Time-horizon doubling periods are often used as a key metric for extrapolating AI capabilities; if growth rates fluctuate substantially depending on domain, baseline, or fitting method, the uncertainty of related predictions will rise significantly.