# OpenAI's agents now finish half its two-to-four-hour research tasks (slim)

> Vlad Cealicu on OpenAI's second research acceleration report. Agents now finish 46% of its two-to-four-hour research tasks without a human, the median researcher runs $2,400 a day of agent time and 27 of 28 researchers would take the latest internal model over a mid-career engineer. The progress is jagged, humans still pick most of the ideas, safety work grew fastest, and OpenAI wants every lab to report the signs of recursive self-improvement.

- Full: https://www.anchorterminal.com/blog/openai-agents-half-the-research.md (~2,700 tokens) · this version ~1,430 tokens · JSON https://www.anchorterminal.com/blog/openai-agents-half-the-research.json · canonical https://www.anchorterminal.com/blog/openai-agents-half-the-research
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-19
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

By Vlad Cealicu*, 19 November 2026. 1,241 words, about 5 min to read.

- 46%: of research tasks estimated at two to four hours of human work that agents finished with no human stepping in, in October, up from 23% in March
- $2,400: a day of agent use at API list prices for the median OpenAI researcher in October, up 62% from September
- 5.7: agent workdays per researcher per day, up from 0.5, an elevenfold rise
- 27 of 28: researchers who would take the latest internal model over a mid-career engineer working with o1-preview

OpenAI, Research acceleration, the jagged frontier [1].

The median researcher at OpenAI ran $2,400 a day of agent time in October, at API list prices, 62% more than in September.

For research tasks OpenAI estimates at two to four hours of human work, the share agents finished without anyone stepping in went from 23% in March to 46% in October.

OpenAI is careful to say it can't fully separate better models from researchers getting better at picking which tasks to hand over, so part of the climb is people learning to delegate [[1]](#ref-1).

## How much they use agents

Every eight hours of agent use counts as one agent workday in OpenAI's numbers.

Priced at API rates, the research organisation's agent use in September came to about twice what it paid its researchers in cash, meaning salary, bonus and benefits.

In a survey of 101 research staff, the median answer to what they'd give up to keep frontier coding agents for three months was 40% of their team's headcount, or 30% of its compute, and 50% of the compute if asked in January 2027. Without agents they reckon a month's work would take three.

On the output side, lines of code changed per active contributor were about 1.3 times the pre-2025 average at the start of 2026 and 6.7 times by 18 October, after a peak of 7.5 times on 7 October.

On OpenAI's own version of the Epoch Capabilities Index its frontier is rising 16.4 points a year since August 2025, though OpenAI notes its models score higher on its own index than on Epoch's.

## Where agents are still bad

OpenAI calls the report the jagged frontier because the agents aren't good at research across the board.

Share of tasks agents finished with no human stepping in, 1 August to 18 October 2026

| Task | Per cent |
| --- | --- |
| Monitoring training runs | 79 |
| Answering a technical question | 72 |
| Launching training runs | 35 |
| Building evaluation and RL environments | 29 |
| Writing inference and serving code | 25 |

Each series has a zero-based scale.

In OpenAI's interviews the researchers say the agents rabbit-hole, write what one of them calls "AI slop" pull requests, make wrong assumptions in analysis and have poor taste in research ideas.

That's the first time I've seen a lab say a model helped with the ideas and not only the code.

## Safety

Safety and alignment saw the biggest absolute rise in documented research effort of any area at OpenAI, up 77% from January to September, as new hires joined and existing researchers moved onto it.

## Publishing the numbers

OpenAI wants every lab to be required to track and publish the signs of recursive self-improvement, and adds that its reporting shouldn't be read as an intention to pursue full RSI.

I like a lab publishing its own acceleration curve with the bad tasks next to the good ones.

OpenAI measuring OpenAI isn't an audit, though.

---

More in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md)

- Older: [Gemini 4 Neon, and more horses in the race](https://www.anchorterminal.com/blog/gemini-4-neon-more-horses-in-the-race.md)
- All posts: https://www.anchorterminal.com/blog/index.md

Anchor Terminal grades the tools AI agents run on, with a source for every score.

## References

Sources read on 19 November 2026.

[1] OpenAI, Research acceleration, the jagged frontier, 18 November 2026. Source for the agent spend per researcher, the chart data read from the page (success by task length by month, lines of code changed per contributor, and OpenAI ECI scores by model), agent workdays, concurrent agents, agent use against non-equity labour costs and the share of research compute, the zero-intervention success rates by task length and by task, the interviews, where the ideas for GPT-6 Astra and GPT-6.2 Astra came from, the surveys, research experiments, OpenAI ECI against Epoch's trend, the growth of safety and alignment work, the call for labs to report indicators of recursive self-improvement and the other labs' reports it cites.

[2] Anthropic, When AI builds itself, undated.

[3] Epoch AI, Trends in AI, updated 5 February 2026. Source for the ECI frontier advancing 14 points a year since the introduction of reasoning models, with a 90% interval of 12 to 17. https://epoch.ai/data/trends

[4] METR, Independent assessment of AI R&D acceleration at Anthropic, 5 November 2026. Source for the high-access evaluation from 12 August to 29 October 2026, Anthropic likely improving its models on well-specified tasks at around twice the March 2024 to October 2025 rate with a range of 1.3 to 3 times, the majority attributed to its use of AI, and Anthropic redacting a majority of the evidence METR found most informative.
