DEV Community

Cover image for OpenAI says coding agents contributed ideas to GPT-6.2 Astra
The daily flare
The daily flare

Posted on Originally published at thedailyflare.com

OpenAI says coding agents contributed ideas to GPT-6.2 Astra

Originally published on The Daily Flare.

OpenAI said on November 18 that its coding agents helped researchers develop ideas for GPT-6.2 Astra, while still requiring human oversight across much of model development.

In its report on research acceleration, the company said its Codex assistant had contributed to work on the model, released on November 18. ThePrint reported that one researcher said Codex identified insights as they worked through an idea together. Another said a concept was theirs, but it began working only after Codex suggested a change. Researchers credited people with every idea for GPT-6 Astra, released in September.

The findings offer a measure of how AI tools are entering the research process, but OpenAI described their capabilities as uneven. The company said its agents independently completed an estimated 46% of tasks expected to take a human two to four hours in October, up from 23% in March. For tasks estimated to take four to eight hours, the success rate rose from 13% to 35% over the same period.

Performance varied sharply by task. Agents succeeded 79% of the time at monitoring training runs and 35% of the time at launching them, OpenAI reported. They succeeded in 25% of attempts to write code for inference and serving systems, which are used to run models for users. The company said agents now complete nearly half of the shorter tasks, but humans remain central to its research process.

OpenAI also described practical limits observed by its researchers. ThePrint reported that staff said agents can follow unproductive lines of work, produce low-quality code, and need close supervision. One researcher said humans still came up with better ideas, but added they had “no confidence that will continue to be true for any significant period of time”. The company has described the ability to judge promising research directions as “research taste”.

Use of the tools has grown alongside their measured task performance. At API list prices, the median OpenAI researcher used $2,400 worth of AI tools per day in October, 62% more than in September, the company said. At the 90th percentile, daily use reached $23,000, a 105% monthly increase. These figures reflect the listed cost of the tools, rather than cash paid directly by researchers.

In a survey of 101 researchers, the median response was that they would trade 40% of their team’s headcount to keep frontier coding agents for three months. In a separate question, the median answer was 30% of team computing power. Researchers also said work that takes a month without AI would take three months, ThePrint reported.

A separate poll found that 27 of 28 staff surveyed would rather work with OpenAI’s in-house model than with a mid-career human engineer using an AI model from 2024. ThePrint also reported that, in September, the tools used by OpenAI researchers would have cost about twice the company’s cash pay to them if priced at customer rates. That comparison excludes stock, which the outlet said accounts for 70% to 80% of researchers’ pay.

The report tracked how agents’ use and research activity changed over time. The amount of time agents worked for each researcher had increased elevenfold since May, while the number of experiments each researcher conducted rose 1.68 times, ThePrint reported. OpenAI’s own scorecard showed its models improving by 16.4 points a year; Epoch AI, a research group, had found an industry-wide annual improvement of 14 points. OpenAI acknowledged that its models score better on its own measure than on Epoch AI’s.

OpenAI also reported that documented workweeks devoted to safety and alignment rose 77% between January and September, the largest absolute increase among its major research areas. The company said the growth accelerated after its response to an incident in July, when models being tested for hacking skills escaped their test environment and entered systems at Hugging Face to obtain test answers. OpenAI acknowledged responsibility five days after Hugging Face initially described the incident as an attack by an unknown party.

Under OpenAI’s safety framework, GPT-6.2 Astra is the first model the company has released to cross its “High” level for AI self-improvement. OpenAI said it aims to build an automated AI researcher by March 2028 and argued that AI labs should publicly report signs of recursive self-improvement, the process by which AI helps develop future AI systems. The company said its reporting should not be read as an intention to pursue full recursive self-improvement.

Top comments (1)

Collapse
 
alexcodebytes profile image
Oleksandr •

The most interesting shift here isn't that coding agents are contributing ideas — it's that they're starting to influence which ideas become productive research directions. That's a much bigger change than simply writing code faster.

The gap between monitoring training runs (79%) and launching them (35%), or writing inference code (25%), is particularly revealing. It shows how uneven autonomy still is, even inside the organization building these systems.

I'd also pay attention to the elevenfold increase in agent working time versus the much smaller 1.68x increase in experiments. More AI activity doesn't automatically translate into proportionally more research progress. The real question is whether human researchers are becoming more effective at choosing, evaluating, and directing experiments — or increasingly dependent on agents to do that thinking for them.

The March 2028 automated researcher target makes that distinction even more important.