Research

OpenAI agents now complete 46% of two-to-four-hour research tasks

OpenAI says coding agents now finish 46% of two-to-four-hour research tasks, while researchers shift toward safety work and report sharply higher spending on agentic tools.

Research acceleration: The jagged frontier
Research acceleration: The jagged frontierblickpixel / pixabay
By Marcus Bennett8 min read

Why it matters

  • By October, agents completed an estimated 46% of two-to-four-hour research tasks, up from 23% in March.
  • Median daily agent spending reached $2,400 in October; the 90th percentile reached $23,000.
  • Daily agent workdays rose from 0.5 to 5.7 per researcher, an elevenfold increase.
  • Safety and Alignment documented workweeks increased 77% from January through September 2026.
  • Logged experiments per active experimenter averaged 1.68 times the 2025 baseline in the four weeks ending October 18.

OpenAI says coding agents completed an estimated 46% of research tasks expected to require two to four hours of human work by October, up from 23% in March. On tasks estimated at four to eight hours, success rose from 13% to 35% over the same March-to-October period.

Building on its September post, OpenAI presented the update as a broader but partial measurement of research acceleration. The company argued that laboratories should track and disclose recursive self-improvement (RSI), meaning AI-assisted acceleration of future AI development, while saying the report did not signal an intention to pursue full RSI.

OpenAI said agents can absorb substantial pieces of AI research, while people remain core to the process. Its task-level results show the jaggedness: agents succeeded 79% of the time when monitoring training runs, but only 35% when launching them and 25% when writing code for inference and serving stacks.

How intensive is agent use?

At API (application programming interface) list prices, the median researcher spent $2,400 per day on agentic tools in October, up 62% from $1,500 in September. The 90th percentile reached $23,000, up 105% from $11,000. OpenAI called the broader rise exponential and linked it to activity, pricier models, model behavior and agent harnesses, the software scaffolding around coding models.

OpenAI defines one agent workday as eight hours of agent use. The daily average rose from 0.5 to 5.7 per researcher, an elevenfold increase; pretraining and posttraining teams averaged roughly eight, while scaling and infrastructure teams averaged roughly four. The interactive measure captures elapsed wall-clock time rather than price or capability and may undercount automations.

By mid-October, more than two-thirds of included researchers reached a daily peak of at least four concurrent agents. More than 10% averaged at least 16 concurrently.

In September 2026, monthly API-equivalent agent spending equaled about twice OpenAI’s non-equity research compensation, with higher ratios in pretraining and posttraining. Equity accounts for 70% to 80% of target compensation. Internal deployment, combining Codex, ChatGPT and API usage, represented only 2% of total research compute.

OpenAI has no agentic coding quotas. A harness efficiency change deployed on August 21 sharply reduced input-token use and recorded spending. Gradual reversion began September 10 for reasons unrelated to spend, after which token spending rose rapidly.

Where is researcher effort moving?

OpenAI classified January through September 2026 contributions from company-public Slack posts, pull requests and internal documents. Classifiers divided each researcher-week equally across verified projects, measuring attention rather than reporting lines; a posttraining researcher splitting time equally between infrastructure and posttraining counts as half a week in each.

Safety and Alignment recorded the largest absolute increase in documented workweeks among OpenAI’s major research areas, up 77% from January to September. OpenAI attributed the near-doubling to new hires and existing researchers shifting toward alignment evaluations, monitorability and reward-hacking research, which examines models exploiting loopholes. Growth accelerated after the July 2026 Hugging Face incident and work on safety cases for frontier AI training.

Every area grew, though OpenAI may revise category boundaries, including the distinction between posttraining and safety. It also frames coding agents alongside data, human labor and compute as inputs to AI research.

Where do agents succeed?

OpenAI measured zero-intervention success with Epoch AI’s O*NET AI R&D taxonomy, an occupation-task framework for AI research. From August 1 through October 18, 2026, success reached:

  • Communicate: 69%
  • Analyze: 58%
  • Build: 43%

Task-level rates diverged: 79% for monitoring runs for issues and 72% for answering technical questions, against 25% for writing inference and serving code and 29% for building evaluation, reinforcement-learning and agent environments.

OpenAI warned that duration-bucket rates reflect capability gains and researchers’ changing delegation choices. The earlier report split usage by output tokens; this update counts how often researchers invoked agents for individual tasks, giving short tasks more weight than token-intensive training runs. Weaker categories could become bottlenecks.

What do researchers say?

From late August to early September 2026, OpenAI surveyed 101 research staff: “Suppose your team’s access to all coding agents better than GPT-5.2 was revoked for the next 3 months. How much would you trade to restore frontier coding agents?”

Median responses showed willingness to trade:

  • 40% of team headcount.
  • 30% of current team compute.
  • 50% of team compute if the question were asked in January 2027, given expected model improvements.

Without frontier coding agents, teams needed a median of three months to produce work equal to their latest month, using only GitHub Copilot and ChatGPT. OpenAI cautioned that self-reports can be inaccurate and its sampling was ad hoc.

In a separate late-October survey of 28 people, 27 chose or leaned toward the latest internal model over a mid-career human research engineer paired with the 2024 o1-preview model. The remaining respondent declined the forced choice and said they would pick GPT-6 Astra.

OpenAI asked one to four researchers who made key contributions in each of three areas—pretraining, posttraining and safety monitoring—to name one to three ideas important to GPT-6 Astra and GPT-6.2 Astra.

It also asked how ideas originated and what role AI played, treating the ability to develop and vet conceptual ideas, or research taste, as a potential RSI bottleneck. Posttraining and safety contributors said people originated every named idea and models helped implement them. Pretraining contributors likewise attributed GPT-6 Astra ideas entirely to people.

For GPT-6.2 Astra, pretraining researchers described developing ideas with Codex. One credited Codex with identifying key insights; another said the main idea was theirs, but began working only after Codex suggested a change.

How are workflows changing?

OpenAI interviewed 10 researchers across pretraining, posttraining and safety. Five themes emerged.

Implementation, debugging and experiment oversight improved. Researcher 7, working on safety evaluations, said:

“I once ported an open-source eval where it was much more involved to get studies and ablations on 'is this a good eval'.”

“It might have been a week and a half or two to really get to the point of ‘Can we use this or not?’ Whereas now, depending on the eval, zero to running it is a day, maybe a few days.”

Research direction remained human-led. Researcher 2, working on reinforcement-learning environments, said:

“The model generally can’t one-shot come up with well-posed problems that it cannot solve itself. And if you ask the model to generate a data point a hundred times, it’ll generate similar data points each time.”

Close review remained necessary because agents launched experiments inefficiently, rabbit-holed, produced “AI slop” pull requests, unusable data artifacts and invalid analysis assumptions. Their weak taste also extended to code, data and writing. Researcher 1, working on posttraining evaluations, said:

“I've tried saying, ‘Just make this task like this other task,’ and been burned too many times when it's not quite right and I need to make edits. Now I say, ‘Go make one. I'll review it and make edits. Then go do the rest.’”

Researchers expected significant improvement in current bottleneck areas. Researcher 8, working on pretraining architecture and data, said:

“I do think I have more taste in terms of the ideas. But I have no confidence that will continue to be true for any significant period of time.”

Researchers also used agents to scan thousands of samples, generate content at scale, diagnose experiments, flag concerning model behavior and build or quality-check datasets. Researcher 3, working on posttraining data quality, said:

“One thing I did recently was look through all of our data… I had it do a couple of rounds and calibrated what it was giving me. … It gave me a really nice, giant sheet with everything organized. … Without AI, this would have been super painful.”

Are capabilities accelerating?

Lines of code changed per active contributor rose sharply during 2026 against the pre-2025 average. OpenAI switched from quarterly averages to a 28-day trailing average, providing a more detailed view, and revised its inclusion criteria to exclude tests. These activity counts do not necessarily imply faster research progress.

Across the four weeks ending October 18, logged experiments per active experimenter averaged 1.68 times the 2025 baseline.

OpenAI added OpenAI ECI (OECI), an internal version of Epoch’s widely used Capabilities Index, as a more direct outcome measure. OECI closely follows Epoch’s methodology but combines public and internal benchmarks. It tracks 20 publicly released OpenAI models and 28 other models.

The OECI frontier for OpenAI models rose 16.4 points per year since August 2025, against Epoch’s 14-point annual trend since late 2024. OpenAI said its models score higher against rivals on OECI than on Epoch’s index, likely because the benchmark mixtures differ. It will keep monitoring OECI and other capability metrics to test whether progress accelerates because of RSI.

What comes next?

OpenAI said these human-AI trends may not predict faster future progress, even as frontier models recently produced sudden mathematical breakthroughs. The company plans to monitor similarly rapid gains in deep learning and refine its methods.

OpenAI plans three additions to its reporting:

  • Differential acceleration of capabilities versus safety research.
  • Leading indicators that detect uplift without completing a full model-development cycle.
  • Capability trends over time and the effect of AI-assisted research on those trends.

OpenAI cited three Anthropic reports on internal deployment-driven acceleration, section 3.4.2 of Google DeepMind’s Gemini 4 Argon Safety and Security Report on the growing share of AI-written code, and a METR collaboration with Anthropic estimating how internal Claude models accelerated the capabilities frontier. Its stated stakes are shared measurement norms, rising competitive pressure, model-development pacing and safeguards that keep people involved as research becomes more autonomous.

OpenAI’s next reporting cycle will test whether early indicators can expose acceleration before a full development cycle ends, while governance keeps people in the loop as models become more autonomous.

Source: OpenAI News

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

791 articles

Related articles

  1. OpenAI researchers say they'd give up 40% of colleagues before their coding agents
  2. OpenAI Predicts AI-Made Discoveries by 2026 as Intelligence Costs Plunge
  3. OpenAI Says Agents Have Replaced Chatbots as Its Default Work Tool
  4. OpenAI's Codex agents now fix ChatGPT performance bugs end to end
  5. OpenAI audit finds ~30% of SWE-Bench Pro coding tasks are broken

« Previous articleNext article »