OpenAI's coding agents can now complete 46 percent of research tasks that would take a person two to four hours — with no human intervention at all. The figure, up from 23 percent in March, comes from "Research acceleration: The jagged frontier," the company's second report on how AI is transforming its own research pipeline, and it arrives with an equally striking price tag: the median OpenAI researcher now runs through roughly $2,400 a day worth of agent compute at API list prices.
The report, published Wednesday, is the most detailed public accounting any AI lab has given of how machine intelligence is reshaping machine intelligence development — the phenomenon researchers call recursive self-improvement (RSI).
The Numbers: Agents Are Getting Faster and More Autonomous
The headline autonomy gains are substantial. For tasks estimated to require two to four hours of human work, zero-intervention success rose from 23 percent in March to 46 percent today. For longer efforts of four to eight hours, success climbed from 13 percent to 35 percent.
OpenAI stresses an important caveat: as agents become more capable, researchers delegate harder tasks within each duration bucket, so the improvement reflects both genuine capability gains and shifting human choices about what to hand off.
The Jagged Part: Performance Varies Wildly by Task
The report's title is no accident. Agents excel at some research work and stumble badly on others. Across task categories measured from August 1 through October 18, 2026, estimated success rates were highest for Communicate tasks (69 percent) and Analyze tasks (58 percent), but dropped to 43 percent for Build tasks.
At the individual task level the spread is even sharper. Agents aced monitoring training runs for issues (79 percent success) and answering technical questions (72 percent), but managed only 25 percent success at writing code for inference and serving stacks, and 29 percent at building evaluation, reinforcement-learning, and agent environments.
OpenAI warns that this heterogeneity matters: as long as model performance across AI R&D tasks is uneven, the weak categories could become bottlenecks that cap overall acceleration. The third-party coverage of the report, including The Next Web, highlighted precisely this gap between headline autonomy numbers and core software engineering work.
What Researchers Say: Three Months of Work in One
The report leans heavily on surveys of OpenAI's own staff. In a survey of 101 research staff conducted in late August and early September, the median researcher said their team would need about three months to complete work equal to their last month of output if they lost access to frontier coding agents and were left with only GitHub Copilot and ChatGPT.
A separate late-October survey posed a striking comparison: 27 of 28 respondents said they expected to be more productive with OpenAI's latest internal model than with help from a mid-career human research engineer paired with a 2024-era model like o1-preview. The lone holdout declined the binary framing — and said they would pick GPT-6 Astra over the human-assisted option anyway.
Humans Still Own the Ideas — For Now
Despite the autonomy gains, OpenAI's analysis of its most important recent research found that human taste still drives breakthroughs. For both the GPT-6 Astra and GPT-6.2 Astra releases, posttraining and safety monitoring researchers said every key idea they named came from people, with models helping implement them.
But the boundary is starting to blur at the edges. For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex — with one crediting the model for identifying key insights as they worked through a problem together, and another saying their main idea only started working well after Codex suggested a change.
Why OpenAI Is Publishing This at All
The company frames the transparency push as a policy argument. OpenAI has called for AI labs to be required to track and publicly report indicators of recursive self-improvement, so that governance debates rest on measurements rather than speculation. The report is explicit that its publication "should not be read as an intention to pursue full RSI," and notes that several other organizations have recently contributed to shared measurement norms for AI R&D acceleration.
It is a notable stance for a competitive lab: publishing internal productivity data invites both scrutiny and imitation. But OpenAI's bet is that industry-wide reporting norms — similar in spirit to compute and safety reporting frameworks — are more valuable than the secrecy alternative, particularly as agents begin contributing to the very systems that will build their successors.
The Bigger Picture
The report lands amid fevered debate about whether AI is already accelerating AI development and what that means for safety and jobs. OpenAI's data adds hard numbers to both sides: agents genuinely substitute for substantial human work at meaningful success rates, yet the most creative and infrastructure-critical tasks remain stubbornly human-dominated.
The jagged frontier, in other words, is real — and the labs racing along it are now measuring their own position in public. For researchers watching from outside, the most important number may not be 46 percent, but the trajectory: every metric OpenAI tracks has roughly doubled since March.
Stay Ahead of AIFollow the latest AI developments in AI research and safety with AI Buzz Wire.
Read more AI news →