Subscribe
AI Models

OpenAI: coding agents now finish 46% of 2–4-hour research tasks unaided

Friday, November 20, 2026 · 5 min read

OpenAI said on November 18 that its coding agents now complete 46% of research tasks estimated to take a human 2–4 hours with zero human intervention, up from 23% in March — and the median researcher now spends $2,400 a day on agent use.

Bullish$MSFT

📊 Scorecard: our Bullish call on $MSFT is scored from $498.30, the last close before it ran. Its first result comes after the next close.

OpenAI: coding agents now finish 46% of 2–4-hour research tasks unaided

The News

OpenAI published "Research acceleration: The jagged frontier" on November 18, its second report on coding agents inside its own research organisation. Agents now complete 46% of research tasks estimated to take a human 2–4 hours with zero human intervention, up from 23% in March. For 4–8-hour tasks the rate rose from 13% to 35%.

Capability is uneven. From August 1 to October 18, agents succeeded 79% of the time monitoring training runs and 72% answering technical questions, but only 35% launching training runs, 29% building evaluation and reinforcement-learning environments and 25% writing inference and serving-stack code.

Agent use keeps climbing. Median researcher daily agent spend hit $2,400 in October, up 62% from September, valued at API rates, the 90th percentile hit $23,000 a day, and agent workdays per researcher rose elevenfold to 5.7 per day from 0.5. OpenAI's own caveat: as agents improved, researchers changed which tasks they delegated, so the gains mix better models with smarter delegation — the report does not separate the two.

Agents finishing nearly half of multi-hour research tasks unaided is a milestone — but the bill, $2,400 a day per researcher, is the other half of the story.

Why It Matters

The numbers matter because they measure something concrete: whether an AI agent can take a job a researcher would spend hours on and finish it without help. In March the answer was yes about a quarter of the time for 2–4-hour tasks; now it is nearly half. For longer 4–8-hour tasks the rate nearly tripled, from 13% to 35%. That is not a marketing claim about capability in general — it is a measured success rate on real work inside OpenAI's own research organisation.

The spending matters because it shows researchers are voting with their budgets. The median researcher spends $2,400 a day on agent use, up 62% in a month, and the top tenth spend $23,000 a day. Agent workdays per researcher rose elevenfold to 5.7 a day — meaning the typical researcher now has the equivalent of nearly six agent workdays running for them every day. People do not spend that kind of money on a tool that is not doing useful work, and that demand is what keeps the computing-power business growing.

The unevenness matters too, because it shows where the frontier actually sits. Agents are strong at monitoring training runs and answering technical questions, and much weaker at launching training runs, building evaluation environments and writing serving-stack code — the jobs where a mistake is expensive and the setup is fiddly. That gap is the honest picture: agents are taking over a growing share of research work, but the hardest, highest-stakes parts still need a human.

Analyst's Take

Our call is Bullish. A doubling in seven months — from 23% to 46% on 2–4-hour tasks, and from 13% to 35% on 4–8-hour tasks — is fast progress on work that used to need a person sitting at the keyboard the whole time. The unevenness is part of the point, not a contradiction: agents are already good at watching training runs (79% success) and answering technical questions (72%), and weaker at the messier jobs like launching runs and writing serving code. That jagged pattern is exactly what you would expect from a technology that is genuinely improving rather than one being oversold.

The spending figures are the other half of the story, and they cut both ways for the market. The median OpenAI researcher now burns $2,400 a day in agent use, up 62% in a single month, and the heaviest users spend $23,000 a day. Agent workdays per researcher rose elevenfold, to 5.7 a day from 0.5. That is real demand for computing power — and for Microsoft, whose cloud runs much of OpenAI's workload, it is revenue. But it is also a cost line growing fast inside the company that builds the models, which is a reminder that the AI boom runs on enormous compute bills.

OpenAI's own caveat deserves weight: as agents improved, researchers changed which tasks they delegated, so the gains mix better models with smarter delegation, and the report does not separate the two. The figures are also OpenAI's self-measured internal estimates, not an independent evaluation. Even so, the direction is clear — agents are doing more of the research work, researchers are spending more to use them, and the jobs they can do unaided keep expanding. Reader poll: Will coding agents replace junior researchers? Vote: Agents replace juniors, Agents assist only, Too early to tell.

Are you bullish or bearish on this story?

By The Numbers

46%Of 2–4-hour research tasks now finished unaided, up from 23% in March
35%Of 4–8-hour tasks finished unaided, up from 13%
$2,400Median researcher daily agent spend in October, up 62% from September
$23,00090th-percentile researcher daily agent spend

FAQ

What is a coding agent?

A coding agent is an AI system that can write, test and run software largely on its own, given a task in plain language. Unlike a chatbot that answers a question, an agent takes a job — such as fixing a bug or setting up an experiment — and works through the steps itself, checking its own results along the way.

What does unaided mean here?

Unaided, or zero human intervention, means the agent finished the whole task without a person stepping in to help, correct it or take over. OpenAI measured the share of tasks its agents completed this way, and that share has roughly doubled since March for tasks estimated at 2–4 hours of human work.

What is the 90th percentile?

The 90th percentile is the level that only the top 10% exceed. When OpenAI says the 90th percentile of researcher agent spend hit $23,000 a day, it means the heaviest-spending tenth of its researchers each burned at least that much on agent use in a day, valued at the prices OpenAI charges outside customers for the same models.

👁

Source: OpenAI

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $MSFT

All $MSFT coverage →

More from AI Agent Update