█

AI/TLDR

OpenAI · 2026-11-18 · major

OpenAI's jagged frontier report — agents now finish 46% of 2–4 hour research tasks

In short

OpenAI's November 18 report says coding agents now finish an estimated 46% of its two-to-four-hour research tasks without help, up from 23% in March. Serving-stack code and training launches remain weak spots.

OpenAI · 2026-11-18 · 3 sources checked

Crypto Briefing header illustration for its report on OpenAI researchers' daily AI token spending
Crypto Briefing

OpenAI's second internal report on AI-assisted research: agent success roughly doubled since March, but the gains are uneven.

Key specs

2–4 hour tasks done alone46% (Oct)
Median daily agent use$2,400 at list prices

Quick facts

PublisherOpenAI, report dated Nov 18, 2026
FollowsSeptember report 'The view inside OpenAI'
4–8 hour tasks done alone35% in October, up from 13% in March
Strongest areaMonitoring training runs: 79%
Weakest areasServing-stack code 25%, launching training runs 35%
Top-decile agent useAbout $23,000 a day at API list prices
DataInternal activity logs, a survey of 101 staff, interviews

What is it?

OpenAI's report 'Research acceleration: The jagged frontier' updates its September look at how coding agents change work inside the lab. It estimates agents finished 46% of two-to-four-hour research tasks without human intervention in October 2026, against 23% in March, and 35% of four-to-eight-hour tasks, against 13%.

How does it work?

The numbers come from internal activity measurements, staff surveys and interviews. Work was attributed by classifying company-visible Slack posts, code-change requests and internal documents. By task type, agents did best at watching training runs for problems (79%) and worst at writing inference and serving code (25%) and launching training runs (35%), measured from August 1 to October 18.

Why does it matter?

The 'jagged' result is the point: OpenAI warns the weak categories could become bottlenecks that cap total research speed-up. Spending is climbing fast too. Median daily agent use hit about $2,400 at list prices in October, up 62% from September. OpenAI also asks labs to publicly track signs of recursive self-improvement, while saying that call is not a plan to pursue it fully.

Who is it for?

AI researchers, engineering leaders and people tracking AI progress

Frequently asked questions

How much do OpenAI researchers spend on coding agents per day?
OpenAI's jagged frontier report puts median daily agent use in October 2026 at about $2,400 per researcher, valued at published API list prices, up 62% from September. The 90th percentile reached about $23,000 a day, up 105%. These are list-price valuations, not OpenAI's internal cash costs, and OpenAI links the growth to more activity and pricier models.
What would OpenAI researchers give up to keep their coding agents?
In a survey of 101 OpenAI research staff from late August to early September, the median answer was 30% of team compute or 40% of team headcount to keep access to agents better than GPT-5.2 over a three-month loss. OpenAI presents these as hypothetical trade-offs in the jagged frontier report, not staffing decisions.
How many agents does an OpenAI researcher run at once?
By mid-October 2026, more than two-thirds of the researchers included in OpenAI's jagged frontier report averaged a daily peak of at least four agents running in parallel, and more than 10% averaged at least sixteen. The September report had counted 3.1 agent workdays for every human workday across OpenAI's research organization.
Are the 46% task-success figures independently checked?
No. The 46% and 35% task-success rates in OpenAI's jagged frontier report are OpenAI's own estimates from internal data. Coverage notes the report does not say how the estimates were calculated or validated, and researchers may simply have improved at choosing which jobs to hand to agents, so the rise may not reflect agent capability alone.

Sources · 3 outlets

Tags

  • openai
  • coding-agents
  • codex
  • ai-research
  • recursive-self-improvement
  • automated-researcher
  • agent-spend
  • ai-for-science
  • research-productivity

← All releases · Learn AI