Researcher looking at a row of jars glowing unevenly, illustrating OpenAI AI agent success rates (AI image)

Illustration

Companies3 min read

OpenAI says agents now finish 46% of its multi-hour research tasks

OpenAI says its AI agents now complete nearly half of the internal research tasks that would take a person two to four hours, without human help: 46% in October, up from 23% in March. The median researcher used about $2,400 of agent time a day at API list prices, according to a report the company published on 18 November.

Key facts

  1. OpenAI published “Research acceleration: The jagged frontier”, a report based on internal measurements, on 18 November 2026.
  2. Agents completed 46% of two-to-four-hour tasks without human help in October, up from 23% in March, OpenAI said.
  3. The median researcher used about $2,400 of agents a day at list prices in October; at the 90th percentile it was $23,000.

Manfred „Luigi“ LugmayrBy Manfred „Luigi“ Lugmayr

Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →

The report, “Research acceleration: The jagged frontier”, is based on internal measurements and follows a first analysis in September. For tasks that would take a person four to eight hours, the zero-intervention success rate rose from 13% to 35%, OpenAI said. The company notes that researchers may now be handing agents different tasks than in the spring, so the rates also reflect changes in which tasks get delegated.

OpenAI agent success rates by task

How well the agents perform depends heavily on the kind of work, OpenAI says, hence the “jagged frontier” in the title. For the period from 1 August to 18 October the report gives these zero-intervention success rates, among others:

TaskSuccess rate
Monitoring training runs for issues79%
Answering technical questions72%
Launching training runs35%
Building evaluation, RL and agent environments29%
Writing code for inference and serving stacks25%

By category, communicating (69%) and analysing (58%) were well ahead of building (43%). Tasks where agents are weaker could become bottlenecks, the company wrote. In interviews with ten researchers, staff said the agents had poor taste in research ideas, code quality and data, went down unnecessary rabbit holes and produced low-quality code changes. One researcher described the workaround: have the agent build a single example, review and fix it, then let the agent do the rest.

The cost: $2,400 per researcher per day

In October the median researcher used about $2,400 of agentic tools a day, up 62% from $1,500 in September, according to the report. At the 90th percentile the figure was $23,000, up from $11,000. Researchers have no internal agent quotas. “Agent workdays”, defined as eight hours of agent runtime, rose from 0.5 to 5.7 per researcher per day. By mid-October more than two-thirds of researchers ran a daily peak of at least four agents at once, and more than 10% ran at least 16.

In September the list-price value of agent use was about twice the research organisation's non-equity pay, OpenAI said; equity typically makes up 70% to 80% of target compensation. The company said internal use accounted for only 2% of research compute. In a survey of 101 research staff from late August to early September, the median respondent said they would give up 40% of team headcount or 30% of team compute to keep frontier coding agents for three months. Experiments per active experimenter recently ran at 1.68 times the 2025 baseline.

OpenAI says it publishes the figures because AI labs should report signs that AI is speeding up the development of future AI. Anthropic and Google DeepMind have released similar measurements, OpenAI noted. Documented workweeks on safety and alignment grew 77% from January to September, OpenAI said, picking up after the Hugging Face incident in July, which also features in California's lawsuit against OpenAI.

Analysis

The rates show where agents hold up today: on tasks with a clear check. A training run either has a problem or it doesn't, and the rate is high. Code for a live product needs a judgement on quality, and there the agent finishes only one task in four on its own. The same pattern appeared on Wednesday in OpenAI's chip design work for AMD, where every change went through fixed verification tools, and in Epoch's test with real office work. For marketing teams, that suggests an order: hand agents campaign monitoring, data checks and reporting first, and keep positioning and sign-off with people. The researcher's method of reviewing one example before scaling could serve as an approval rule for marketing teams.

The cost figure should be read with caution. OpenAI prices usage at its own list rates, which it does not pay internally, and sets no quotas. Outside customers pay for their usage, as Microsoft's cost study set out this week. Teams rolling out agents widely should put usage per person next to staff costs each month and record, per task, how often the output was usable without rework. Only that rate shows whether an agent takes work off people or just moves it. The guide AI agents for marketing covers where to start.

  • Share

Got AI news? Submit news →

FAQ

How much do OpenAI researchers spend on AI agents?

According to OpenAI's 18 November 2026 report, the median researcher used about $2,400 of agents a day at API list prices in October, and the 90th percentile $23,000. There are no internal quotas.

Which tasks do OpenAI's agents handle best?

Monitoring training runs (79% without human help) and answering technical questions (72%) scored highest; writing code for inference and serving stacks scored lowest at 25%.

Source

OpenAI: Research acceleration: The jagged frontier

More sources

  • OpenAI – Erster Bericht vom September 2026, nur als Hintergrund.

Black Friday 2026

Black Friday deals on AI tools, straight to your inbox.

I collect the AI tool deals worth having and send them in the AI Briefing, along with the AI news that matters, every Wednesday and Friday. Five minutes to read.

  • Black Friday deals on AI tools, collected for you
  • The key AI news, briefly explained
  • Wednesday and Friday, unsubscribe in one click

In the AI Briefing right now

Image for the story: ElevenLabs takes 55% off annual Creator and Pro plans for Black FridayDealElevenLabs takes 55% off annual Creator and Pro plans for Black Friday
Image for the story: Hackers hijack Microsoft's Agent Framework on PyPI to steal cloud keysNewsHackers hijack Microsoft's Agent Framework on PyPI to steal cloud keys
Image for the story: OpenAI says agents now finish 46% of its multi-hour research tasksNewsOpenAI says agents now finish 46% of its multi-hour research tasks
Blonde woman with glasses smiling as she scrolls on her phone in an office

Manfred “Luigi” LugmayrFrom Manfred “Luigi” Lugmayr: 30 years in marketing, seven of them as CMO at burgerme.