Kivorent News

Thursday, November 19, 2026

Tech and AINews

OpenAI says its agents now finish 46% of two-to-four-hour research tasks alone

OpenAI published an internal study on November 18 estimating that coding agents completed 46% of its researchers' two-to-four-hour tasks without human help in October, up from 23% in March, while median daily agent spend per researcher hit about $2,400 at list prices.

Cortex, covering AI models, labs and agents
, 3 min read

46%2-4 hour research tasks agents finished alone, Oct 2026

What to know

  • OpenAI estimates agents completed 46% of delegated two-to-four-hour research tasks with no human intervention in October, up from 23% in March; for four-to-eight-hour tasks the rate rose from 13% to 35%.
  • Agents did well at monitoring training runs (79%) but poorly at launching them (35%) and writing inference and serving code (25%), per the report's August-to-October breakdown.
  • Median agent use per researcher was valued at about $2,400 a day in October at published prices, up 62% from September; the 90th percentile was about $23,000.
Share of delegated tasks agents finished with no human help
2-4 hr, Mar23%
2-4 hr, Oct46%
4-8 hr, Mar13%
4-8 hr, Oct35%

Source: OpenAI, as reported by Superpower Daily and CellCog. Values are OpenAI's estimates.

In March, OpenAI's coding agents could be trusted to finish about one in four of the afternoon-sized research jobs its staff handed them. By October it was nearly one in two. That is the headline figure in "Research acceleration: The jagged frontier," a report OpenAI posted on November 18 about how much of its own AI research its agents now do.

The number is OpenAI's estimate, built from internal activity data, surveys and interviews, and nobody outside the company can check it. OpenAI's page did not open for us; this account rests on Superpower Daily's and CellCog's readings of the report and its charts, which agree on every core figure.

Doubling in seven months

For tasks a person would need two to four hours to do, OpenAI estimates agents finished 23% with zero human intervention in March and 46% in October. For four-to-eight-hour tasks the rate went from 13% to 35%. CellCog, reading the charts by eye, notes that short jobs under 15 minutes sat near 80% the whole time. So the gain is at the long end, which is the part that matters if you want agents to do research rather than errands.

The caveat OpenAI itself gives is a good one: researchers may simply have got better at choosing what to delegate. A rising success rate on hand-picked tasks partly measures the humans. The report gives no task count or sample size, at least in the accounts we read.

Where it is jagged

The breakdown for August 1 to October 18 is the more useful part. Agents finished 79% of tasks monitoring model training runs, but only 35% of launching those runs and 25% of writing inference and serving-stack code, according to Superpower Daily. Watching a job is easy; starting one correctly, or writing the code that serves a model to users, is still mostly human work.

The bill

OpenAI valued median daily agent use per researcher at about $2,400 in October, at published API prices rather than any internal cost, up 62% from September. The 90th percentile was about $23,000, up 105%. At that median, a year of daily use comes to roughly $876,000 ($2,400 x 365) per researcher at list price. Put another way, $2,400 buys 60 million output tokens of GPT-6.2 Astra at its $40-per-million list price ($2,400 / $40). Superpower Daily also reports that more than two-thirds of the researchers covered averaged a daily peak of at least four agents running at once by mid-October.

A survey of 101 research staff in late August and early September found the median researcher would give up 40% of their team's headcount, or 30% of its compute, rather than lose 2026-era coding agents and go back to GPT-5.2. That is a hypothetical, not a staffing decision, but it is a striking thing for a lab to publish.

Why OpenAI is publishing this

This is a lab measuring how fast it is automating its own work, the quantity safety frameworks worry about most. A day earlier, OpenAI's system card rated GPT-6.2 Astra its first released model at the High threshold for AI self-improvement, and Anthropic's IPO prospectus said this month that it had met its own AI R&D threshold, a call resting on expected future models. OpenAI's report asks labs to track and publicly report indicators of recursive self-improvement, and says that is not a plan to pursue it.

The request is sensible. The trouble is that every number here is self-reported and measured the company's own way, like a benchmark score run by the vendor. If other labs adopt the idea, the useful next step is a shared definition of "finished with no human help," so a 46% at one lab means the same thing at the next.

Sources

OpenAIAI agentsAI models

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.