Intel / article

Research acceleration: The view inside OpenAI

Source: simonwillison.net
Date: Published Sep 6, 2026

From the Daily Brief

A recent post examines how OpenAI's own research team relies on coding agents day to day, using an internal chart of AI spend per researcher as its anchor. That chart shows a sharp jump in spend during late July, which lines up with when the model later released publicly as GPT-6 Astra reportedly reached internal users first. The timing suggests that early access to a stronger model changes how much researchers are willing to spend running agents against real problems, not just how they talk about the tool. It also gives outside teams a rare, concrete number to compare against instead of vague claims about internal adoption. For teams building their own coding agents, the useful signal is less about the specific dollar figure and more about the shape of the curve. Spend rose before the public release, suggesting the biggest gains from a new model show up internally well ahead of launch day.

Read the 2026-09-07 Brief →

Context

OpenAI reports that by mid-August 2026, its median researcher used more than $600 per day of coding- valued at API prices. That figure measures consumption using public rates, rather than the company's actual computing bill. The report defines researchers broadly, including staff who build infrastructure and support research projects.

The distinction matters when judging research acceleration: usage and generated code are easier to count than scientific progress. OpenAI reports more experiments per active experimenter alongside greater Codex adoption, but also says its available compute grew significantly since 2025. That makes it difficult to isolate the agents' contribution. The company presents these as preliminary measurements, with people still setting priorities and deciding which results to pursue.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source simonwillison.net