OpenAI says coding agents are speeding up its own research

OpenAI published a look at how coding agents are changing work inside the lab, backed by internal usage data rather than the usual anecdotes. The headline claim is that the number of experiments per active researcher kept climbing through 2026 and hit an all-time high in August, the highest since the company started tracking this in January 2025. OpenAI ties the rise to growing adoption of Codex, its coding agent, though it is careful to note that available compute also grew over the same period, so the two effects are hard to fully separate.

Two numbers stand out. Over the past six months, OpenAI says the share of its research compute spent on internal coding inference grew roughly 100-fold, and internal agentic token usage rose about 22-fold. Researchers now run agents in concurrent sessions through the day, shipping code faster and launching more experiments, with agents handling harder tasks and succeeding more often. This is self-reported data from a company with an obvious interest in the story it tells, so treat the exact multiples with some caution. Still, it is one of the more concrete public accounts of AI tools measurably changing the pace of frontier research from the inside.

Why it matters

If you are trying to judge whether AI is starting to accelerate its own development, this is real internal evidence to weigh, not speculation. For engineering leaders, the pattern is a signal about where coding-agent throughput is heading, and a reminder to measure experiments shipped, not just tokens spent, when you evaluate the same tools on your team.