OpenAI’s Research Agents Now Cost 2x Researchers’ Cash Pay

OpenAI says the agents its researchers use now cost about twice what it pays those researchers in cash. In September, the API-equivalent value of agent use was “about twice its non-equity compensation”, per a research report it published on 18 November.

That excludes equity, which OpenAI says is 70 to 80% of researchers’ target pay.

Business Pill: The Cost of Checking

An evergreen explainer on who pays to check work, and why cheap verification changes what can be automated. It does not discuss OpenAI’s report.

The key insight: OpenAI says the API-equivalent value of its researchers’ agent use was about twice their non-equity pay in September, with median use of about $2,400 a day in October.

What OpenAI Measured

At API list prices, the median researcher used about $2,400 a day of agentic AI tools in October. The 90th-percentile researcher used about $23,000.

Those figures were $1,500 and $11,000 in September, OpenAI says, up 62% and 105%.

Agent workdays rose to 5.7 per researcher per day from 0.5, an elevenfold increase. OpenAI counts every eight hours of agent use as one workday.

OpenAI chart: ratio of agentic token costs to non-equity labor costs by research area, January to September 2026; overall 2.0x in September. Source: OpenAI.
OpenAI’s own chart: agent spend as a multiple of researchers’ non-equity labor costs, by research area. Source: OpenAI, Research acceleration: The jagged frontier, 18 Nov 2026.

Internal deployment still takes only 2% of total research compute, the report says.

OpenAI's estimated success rate for agents finishing research tasks without human help, March vs October 2026,
OpenAI’s estimated success rate for agents finishing research tasks without human help, March vs October 2026, by how long the task would take a person. Data from OpenAI’s report.

What Was Posted on X

OpenAI announced the report on X at 23:07 UTC on 18 November, saying researchers’ AI use continues to rise exponentially and that labs should be required to track and publicly report indicators of recursive self-improvement. This piece is built from the report the post links.

Andrew Curran posted at 00:00 UTC on 19 November that the 90th-percentile OpenAI researcher now uses around $23,000 worth of tokens per day, up 105% from September. The same figures appear in OpenAI’s report, which this piece is built from.

How Much Agents Now Finish Alone

From March to October, agents’ estimated success without human help rose to 46% from 23% on tasks of two to four human hours.

On four-to-eight-hour tasks it rose to 35% from 13%. Success is uneven: 79% for monitoring training runs, 25% for writing inference and serving code.

In a survey of 101 research staff, the median answer was that they would trade 40% of team headcount, or 30% of team compute, to keep frontier coding agents for three months.

Update, 19 November 2026, 00:19 UTC: OpenAI puts its own capability-index trend beside Epoch AI’s. In the same report OpenAI describes OpenAI ECI (OECI) as “an OpenAI-specific internal version of the Epoch Capabilities Index”, built with a method “very similar to Epoch’s methodology” but on “a combination of public benchmarks and OpenAI internal benchmarks” (OpenAI). OpenAI writes: “We find that the OECI frontier for OpenAI models has increased by 16.4 points per year since August 2025, compared to a trend that Epoch observed of 14 points per year” since late 2024. It adds that “OpenAI models score higher relative to competitors on OECI than they do on Epoch’s ECI, likely due to differences in the benchmark mixture”.

Why OpenAI Published It

OpenAI says AI labs should be required to report signs of recursive self-improvement (RSI), where AI speeds up the building of future AI.

It adds that its reporting “should not be read as an intention to pursue full RSI”.

For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex, the report says. Human effort on safety and alignment rose 77% from January to September.

Connecting the Coverage

The report lands on a day of AI-builds-AI news: Vals AI’s account of Astra’s Lean proof and OpenAI’s AMD chip-design results. We also covered Mark Chen on self-improving AI and Brockman’s expanded remit.

Critics of the race cite the same signal: see the MisAnthropic Statement and Dan Hendrycks’ exit from advising AI firms.

The Structural Read

As we read the report, the cost line has flipped: at list prices, agents now outspend the cash payroll they assist, while using a sliver of research compute.

As we read the success rates, agents are taking over the checking work first, such as monitoring runs, and lag on building new systems, such as serving code.

OpenAI, Research acceleration: The jagged frontier, 18 Nov 2026

“we believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI)”

Three Implications

FOR AI BUYERS OpenAI says median agent use reached about $2,400 per researcher per day in October, at API list prices.

FOR RESEARCH MANAGERS OpenAI reports agents finished 46% of two-to-four-hour tasks without help in October, up from 23% in March.

FOR POLICY READERS OpenAI says labs should be required to track and publicly report indicators of recursive self-improvement.

What Is Not Established

The spend figures are at API list prices, not OpenAI’s own cost. The labor comparison leaves out equity.

If equity is 70 to 80% of target pay, cash pay is roughly a fifth to a third of it. That would put agent spend near half of full target pay (our rough arithmetic; OpenAI’s labor base also includes benefits and taxes).

Success rates are OpenAI’s estimates, and it says researchers may have changed which tasks they hand to agents. We did not contact OpenAI.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

A Business Engineer analysis of the fight to own the software layer that runs AI agents.

Read the Map of AI →

The Bottom Line

By OpenAI’s own count, agents now cost its research arm about twice its cash payroll, while finishing 46% of two-to-four-hour tasks alone and using 2% of research compute.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every figure, quote and chart value is from OpenAI’s 18 November report ‘Research acceleration: The jagged frontier’ and OpenAI’s post on X. The chart image is OpenAI’s own, credited. The share-of-full-pay estimate is our rough arithmetic on OpenAI’s 70 to 80% equity figure. We did not contact OpenAI.

Sources: OpenAI, ‘Research acceleration: The jagged frontier’, research report, 18 Nov 2026 · OpenAI on X, 18 Nov 2026, 23:07 UTC

Scroll to Top
FourWeekMBA