AI Building AI: OpenAI’s Agents Finish Half the Research

Human Verified

Share to: , and

A researcher silhouetted before glowing monitors showing abstract charts and network graphics in a dark lab at night

OpenAI says its coding agents now finish nearly half of its two-to-four-hour research tasks without human help. The company published the numbers on 18 November in a report on research acceleration. Success rose from 23% in March to 46% in October. Humans, it stresses, remain central.

Where Agents Win — and Where They Stall

The gains are uneven, and OpenAI does not hide it. Agents succeeded without help in 79% of cases when monitoring training runs, but only 35% when launching them. Writing code for inference and serving stacks came in at 25%. Communication tasks reached 69%, analysis 58% and building 43%.

Longer work is harder. For tasks estimated at four to eight hours, success without intervention rose from 13% in March to 35% in October. OpenAI cautions that researchers changed what they delegate as agents improved, so the figures blend capability gains with shifting choices.

The Bill Keeps Climbing

Usage is surging faster than the success rates. At list prices, the median researcher now spends about $2,400 a day on agentic tools, up 62% since September. The heaviest users, at the 90th percentile, reached $23,000 a day, up 105%. Each researcher now averages 5.7 agent workdays per day.

Agents now run in parallel, not one at a time. By mid-October, more than two-thirds of researchers averaged a daily peak of at least four agents working at once, and over a tenth averaged sixteen or more. The value of agent use now runs at about twice researchers’ non-equity pay.

Humans Still Hold the Ideas

The big ideas still come from people. For the GPT-6 Astra release, researchers credited every idea to humans. For GPT-6.2 Astra, Codex helped develop ideas in conversation, once spotting a key insight. Surveyed staff would trade 40% of team headcount to keep frontier agents for three months.

Safety work is growing fastest inside the lab. Documented Safety and Alignment workweeks rose 77% from January to September, the largest increase of any research area. Growth quickened after the July Hugging Face incident, as OpenAI built safety cases for frontier training. Watching how models think is part of it.

What the Outputs Show

Output measures point the same way, with limits. Logged experiments per researcher ran 1.68 times the 2025 baseline in the four weeks to 18 October. OpenAI’s capability index rose 16.4 points a year since August 2025, faster than Epoch’s 14-point trend. Lines of code changed per contributor also climbed sharply.

Our Read

The numbers matter less than the fact OpenAI published them. No lab has put this much of its own acceleration on the record, and the report asks rivals to publish the same indicators. That turns recursive self-improvement from a warning into something measurable, and therefore governable.

The bottleneck is the real story. Agents already handle monitoring and answering questions well, but building and launching lag far behind. Until the Build category catches up, the speed-up stays capped at the weakest link. Watch the cost curve: at $23,000 a day for heavy users, acceleration is getting expensive.

Why It Matters

This report sets the terms for a debate that is only starting. Regulators and rivals now have a yardstick for how fast AI is improving itself, and OpenAI has staked its position on transparency. If other labs follow, the industry gains an early-warning system. If not, its numbers stand alone.

Background

OpenAI first reported on research acceleration in September, when agents delivered 3.1 workdays per researcher per day. This update extends that picture with task-level success rates and spending data. The company says it will keep refining the measures and publish more, including leading indicators of acceleration.

Source: OpenAI’s research report

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.