On November 18, 2026, OpenAI published new data on internal research acceleration on its official blog: from March to October, coding agents' success rate at completing two-to-four-hour research tasks on their own rose from 23% to 46%, and for four-to-eight-hour tasks from 13% to 35%. The gains are real, but uneven — OpenAI's own phrase is "the jagged frontier."
Agents Are Being Used More Heavily — and Their Weak Spots Are Clear
The type of task decides success or failure. Monitoring training runs succeeds 79% of the time and answering technical questions 72%, while writing code for inference and serving stacks manages only 25%, and building evaluation and reinforcement-learning environments 29%. Researchers' daily agent spending is also climbing fast: at API list prices, the median reached $2,400 in October, up 62% from September, while the top 10% of researchers used $23,000 a day. Asked what they would trade for three months of access to frontier coding agents, the median researcher answered 30% of team compute — or 40% of team headcount.
Safety Spending Is Growing Fastest — and Humans Stay in the Loop
Notably, the fastest-growing area of investment in 2026 was not pretraining or posttraining but safety and alignment: related research hours rose 77% from January to September, accelerating further after the Hugging Face incident in July. The report also traces where the key ideas behind the GPT-6 Astra and GPT-6.2 Astra releases came from. OpenAI's conclusion is restrained: agents are genuinely rewriting the research process, but humans remain central — where ideas come from, which tasks are worth handing to agents, and who is responsible when things go wrong still depend on researchers themselves. The report's subtext is also worth noting: labs should publicly track and report such "recursive self-improvement" indicators, or the outside world will have no way to judge how fast AI is accelerating research.