AI Research

OpenAI agents do busywork, not research

OpenAI says agents finish 46% of short research tasks alone, but build work still lags badly.

StartupHub.ai Staff

By StartupHub.ai Staff, reviewed by Daniel Singer, Editor

3 min read

OpenAI research chart showing agent task success rates
Credit: OpenAI News

OpenAI says its coding agents now finish 46% of two to four hour research tasks with no human help.

The obvious read is that recursive self improvement has begun, with agents doing real AI research while humans watch, and I find that claim premature because the measurement moved as fast as the models.

Busy is not autonomous.

OpenAI reports success rose from 23% to 46% on two to four hour tasks from March to October, the company says, and from 13% to 35% on four to eight hour tasks. OpenAI also admits it updated the method and that researchers choose harder or easier tasks as agents improve.

That selection effect matters. Success hits 79% for monitoring training runs, OpenAI says, but drops to 35% for launching runs and 25% for writing inference and serving code. Communicate tasks score 69% while Build tasks score 43%.

For that obvious reading to hold, strength in monitoring and Q&A would have to transfer to the bottlenecks that actually create capability. It has not. OpenAI calls the uplift jagged, and low scoring build and eval work can still gate the whole loop.

Usage is surging even where autonomy lags. Median daily agent spend reached $2,400 per researcher at API list prices in October, OpenAI says, up 62% from September, while 90th percentile use hit $23,000, up 105%. Agent workdays rose elevenfold to 5.7 per researcher, with pretraining and posttraining near eight per day.

List price math flatters the story. Internal agent value now equals about twice non-equity pay, OpenAI says, but equity is 70% to 80% of researcher target pay and is excluded. Internal agent use is still only 2% of total research compute.

The strongest fact for OpenAI comes from its own staff. In a survey of 101 researchers, OpenAI says the median respondent would trade 30% of team compute or 40% of headcount to keep frontier coding agents for three months. That is a real revealed preference, even with a small ad hoc sample.

Competition sharpens the skepticism. Anthropic says Claude Fable 5.2 can run ambiguous engineering work for days and coordinate hundreds or more than a thousand subagents. OpenAI reports more than two thirds of researchers peak at four or more concurrent agents, with just over 10% at sixteen or more.

Safety work is growing alongside agent work. Documented Safety and Alignment workweeks rose 77% from January to September, OpenAI says, with a clear lift after the Hugging Face incident in July. That fits a wider containment push, not a hands off lab.

If you are building a startup, rent the agents for throughput and keep humans on launches, serving code and evals.

Mentioned in this story

Sources

Get the free daily brief every morning

Every morning the day's AI news, new AI startups and new funding rounds. Free and takes a few minutes to read, unsubscribe at any time.

Join 64,000 other subscribers.

TopicsOpenAIAnthropicClaude Fable 5.2CodexHugging FaceGPT-5.2

StartupHub.ai StaffDaniel SingerWritten by StartupHub.ai Staff, Editorial teamReviewed by Daniel Singer, EditorAbout

StartupHub.ai Staff, Editorial team

The staff writers of StartupHub.ai, ranging from investment analysts to avid AI tool users, early adopters and critical enthusiasts. Backgrounds span engineering, business and the arts. We hold every piece to rigorous standards of research and review.

Daniel Singer, Editor

Daniel Singer is an AI expert who has been writing about, covering and building AI technology since 2018.

More from Daniel Singer

Test your Agent Readiness now

Paste your website. How well ChatGPT, Claude and the other agents can read and use a site, in about a minute.

Watch

The videos behind the news, one flip at a time

The feed mixes the videos from our articles with startup demos and you flip through them like TV channels for free with no signup.