AI Briefing
KOSign in

Analysis of OpenAI data suggests AI R&D time horizons may double every 2 months, outpacing METR benchmarks

·2026.11.09 09:00

Key point

Analysis of OpenAI's public data suggests internal AI R&D time horizons may be growing faster than METR benchmarks, with estimates ranging from 2 to 3 months depending on the statistical method used.

1 / 10

Details

  • OpenAI internal AI R&D time horizons show varying doubling times depending on the statistical method: linear interpolation suggests a 68-day doubling time for 50% accuracy tasks, while fitting a logistic or Weibull distribution yields an estimate of approximately 2 months.
  • This potential 2-month doubling rate is significantly faster than the 4-7 month doubling time observed in METR software task benchmarks.
  • At 80% accuracy, OpenAI's time horizon was approximately 15 minutes as of July, a level comparable to GDM's internal AI R&D metrics but lower than METR headline figures.
  • The author notes that the rapid growth rate is sensitive to specification; under linear interpolation, 80% reliability is described as rising very slowly, whereas under logistic/Weibull fits, it is described as rising very quickly.
  • The low absolute level of time horizons (15-20 minutes) may result from measuring high-context human baselines, where METR time horizons are roughly 10x lower than low-context baselines, or from the inherent difficulty of the AI R&D domain.
  • The ratio of 80% to 50% time horizons at OpenAI is approximately 15x, compared to a 4x ratio in METR data, a discrepancy the author attributes to potential differences in task context or domain-specific variance.
openaibenchmarkingmetranalysisai-rndtime-horizons
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

TLDR AIMETR Report Finds AI R&D Acceleration at 2x Speed, Experts Suggest RSI Could Occur Within a YearTLDR AIMETR Report Suggests Anthropic Has Crossed 'Dramatic Acceleration' ThresholdOpenAI BlogOpenAI Registers Business Ownership in ChatGPT and Launches Agent Features in BetaRedditClaude reaches 61% of AI builder job postings, tying with OpenAI within margin of errorHacker NewsNvidia Used Circular Financing to Keep GPU Prices High, WSJ Report ClaimsHacker NewsOpenAI Disbands Robotics Team, Pivots to General Models for Embodied AI