AI coding agents finish 46% of OpenAI research tasks

AI coding agents accelerating research work at OpenAI

AI coding agents now complete 46% of research tasks estimated at two to four hours of human work, with no human help. In March the figure was 23%. OpenAI shared the numbers in an update on how coding agents are changing its own AI development, based on internal measurements.

Success on longer tasks is rising as well. For tasks estimated at four to eight hours, the rate went from 13% in March to 35% in October.

Uneven gains across task types

OpenAI calls the uplift jagged. A task-level breakdown shows how much the kind of work matters:

  • 79% autonomous success for monitoring training runs
  • 35% for launching training runs
  • 25% for writing code for inference and serving stacks

The company adds a caution. Researchers may have changed which tasks they hand to agents, so shifts in these rates reflect both better capability and different delegation choices. Humans remain a core part of the research loop.

Spending is climbing fast

Priced at API list rates, the median researcher spent $2,400 a day on agents in October, 62% more than September’s $1,500. The 90th percentile reached $23,000 a day, up 105% from $11,000. OpenAI links the growth to more activity, pricier models and changes in model behavior and agent tooling.

Measured in time, agent workdays per researcher per day (eight hours of agent use counts as one) grew from 0.5 to 5.7, an elevenfold increase. Pretraining and posttraining teams averaged roughly eight workdays per researcher each calendar day, and scaling and infrastructure about four. The metric mostly counts interactive use, so automations may be undercounted.

Concurrency is up too. By mid-October more than two-thirds of researchers in the sample averaged a daily peak of at least four agents at once, and over 10% averaged at least sixteen.

Agents versus pay

As of September 2026, the API-equivalent value of agent use in OpenAI’s research organization was about twice researchers’ non-equity compensation. Equity typically makes up 70 to 80% of target pay, so the ratio is a partial view. Pretraining and posttraining ran higher than other groups.

Even so, internal deployment, which covers internal Codex, ChatGPT and API use, accounts for only 2% of total research compute.

A survey adds a sense of value. Researchers were asked what they would give up to keep frontier coding agents for three months. The median answer was 30% of team compute in one question and 40% of team headcount in another.

More time goes to safety work

OpenAI tracked where researchers spent their time from January through September, using company-public Slack posts, pull requests and internal documents. Safety and Alignment saw the largest absolute increase of any research area. Its documented workweeks rose 77% between January and September, from new hires and from researchers moving over.

That area covers alignment evals, monitorability and reward hacking research. Growth picked up after OpenAI’s response to the Hugging Face incident in July 2026 and its later work on safety cases for frontier AI training.

The company notes the line between posttraining and safety work is not always clean and it may revise the definitions.

Why OpenAI is publishing this

OpenAI wants AI labs to be required to track and publicly report signs of recursive self-improvement, where AI speeds up the building of later AI. It says openness is needed so the public can debate capabilities, risks and safeguards, and that the report does not signal a plan to pursue full recursive self-improvement. Other organizations have recently published similar measurements.

The report follows a September post on the same topic. It sorts findings into inputs (coding agents, data, labor, compute), process, and outputs in the form of model capabilities.

What to watch in the next report

The next update will show whether the 46% rate for two-to-four-hour tasks keeps climbing and whether the weak spots, such as serving-stack code at 25%, close. Spending growth is the other number to track: a median of $2,400 a day per researcher is already far above September. OpenAI describes its picture as broader but still partial.

Source: OpenAI

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.