OpenAI Unveils AI Research Acceleration Metrics: A Three-Layer Measurement Framework

OpenAI's Internal AI Research Acceleration Methodology

OpenAI has published internal data addressing a previously speculative question: to what extent do AI agents accelerate AI research itself? The most valuable contribution is not the headline numbers but the first public disclosure of how to measure research acceleration—particularly how the tasks researchers delegate to agents are evolving.

Three-Layer Measurement Framework

OpenAI defines a progression: AI Research Intern (achieved September 2026) — systems that complete defined research tasks under human guidance, including tasks taking skilled researchers days; next target: Automated AI Researcher by March 2028. To track progress, OpenAI employs three measurement layers:

Usage — how heavily agents are used (token consumption).

Tasks — what work agents actually do (core of the analysis).

Results — how many tasks are completed successfully.

Usage Layer: Token Consumption Metrics

Median researcher consumes over $600/day in tokens (API-equivalent pricing); top 10% exceed $7,000/day . For every human workday, the organization now logs 3.1 agent workdays — up from less than 1 before June 2026. OpenAI acknowledges usage numbers are "easy to measure, hard to interpret": high token burn does not guarantee faster research.

Task Layer: Structural Shift in Delegation

Using a taxonomy adapted from Epoch AI (inspired by O*NET), OpenAI classifies agent token usage across six R&D stages:

Decide : what to pursue, what to drop, compute allocation.

Design : research ideas and engineering approaches.

Build : code and datasets.

Run : training, evaluation, hardware, deployment.

Analyze : experiment, model, and deployment results.

Communicate : reporting conclusions, feedback, status, decisions.

From January to August 2026, all six categories grew — agents now permeate every R&D phase, not just coding. However, growth is uneven: research infrastructure code remains the largest share, while technical support and run-monitoring categories rose sharply . High-level planning (Decide) remains a tiny fraction — humans still control research direction.

Two concrete signals reinforce this structural change:

Agents excel at debugging internal research infrastructure failures, a genuine bottleneck.

Researcher office-hours attendance dropped continuously in 2026; one team cancelled office hours entirely to reassign staff to system improvements.

Daily posts in the main internal technical-support channel fell, and OpenAI confirms these questions did not migrate to other human-support channels — demand was absorbed by agents.

Results Layer: Success Rates and Human Intervention

Using an internal agentic classifier, OpenAI scored tasks with verifiable outcomes, bucketed by estimated human time. From January to July, success rates rose across all difficulty buckets , including tasks requiring several human hours. Yet, over half of successful 4–8 hour tasks required at least one human intervention — agents handle longer tasks but still need human guidance on complex work.

Key Metrics Summary

Usage : Median researcher daily token cost — > $600

Usage : Agent workdays per human workday — 3.1

Task : Agent usage across six R&D categories — All increased

Task : High-level planning share of agent output — Negligible

Result : Successful 4–8 hour tasks needing human intervention — > 50%

Public Disclosure as Methodology

The publication's logic: set dated targets (Intern achieved, Researcher slated for March 2028), then openly report the three-layer measurement — usage, tasks, results — including numbers, failures, and acknowledged limitations. OpenAI commits to slowing or halting if unmanageable safety risks emerge and calls for industry-wide public tracking of recursive self-improvement (RSI) progress.

Related Resources

OpenAI blog:

https://openai.com/index/research-acceleration-view-inside-openai/

Navier-Stokes solution (165-page paper): https://openai.com/index/navier-stokes-solution/ and

https://cdn.openai.com/pdf/32d9f210-8b73-45e0-91bc-82a30aef8a9a/navier-stokes.pdf