Model

OpenAI's agent bill now runs at twice its researchers' cash pay

OpenAI's own report prices agent use at twice its researchers' cash pay while internal deployment is 2 percent of research compute, and the number lands the week Anthropic lists and OpenAI settles for half its raise. The bill stacks inference on the chip budget, and holders pay it.

Vincent JiangVincent Jiangthree-minute read
Share
Sam Altman seated at a long table among technology executives at a White House dinner
Sam Altman, right of center, at the White House dinner for business and technology leaders on September 4, 2025. OpenAI's own report now prices its researchers' agent use at about twice their non-equity pay.

OpenAI$1.18T — OpenAI, private, latest valuation $1.18T has put its own research loop on the record, and the invoice is the news. A report published on November 18, "Research acceleration: The jagged frontier," counts what the lab's coding agents cost and do 1.

The median researcher burned $2,400 of tokens a day in October at API list prices, up 62% in a month; the 90th-percentile researcher burned $23,000, up 105% from $11,000 1. A heavy user's habit annualizes to roughly $8 million at public rates 1. The median answer across 101 researchers surveyed: they would trade 40% of team headcount, or 30% of team compute, to keep the agents for three months 1.

One line prices the whole trade

The report's quietest line is the one that prices the whole trade: as of September 2026, the API-equivalent value of agent use across OpenAI's research organization was about twice the researchers' non-equity compensation, and equity is 70 to 80% of target pay 1. If equity is three-quarters of the package, cash pay is roughly a quarter of it, so the agent bill works out near half of all-in research wages. Those dollar figures are list prices, what an outside customer would pay, so they mark a ceiling on OpenAI's actual cost; the true number is not disclosed 12.

A labor line, not a chip line

Yet this spend does not compete with the chips investors finance. Internal use of Codex, ChatGPT and the API accounts for just 2% of total research compute 1. Agent research is inference stacked on training. That is the demand picture Nvidia−3.49% — Nvidia, down 3.49 percent today described on its November 18 call: OpenRouter token volume up 35-fold year over year, agentic workloads about 80% of tokens, token prices down about 60% since May while GPU rental rates rose 4.

The heaviest users more than doubled their daily agent bill in a month

  • Median researcher
  • 90th-percentile researcher
$0$10,000$20,000$30,000September 2026October 2026$23,000
Data
Median researcher90th-percentile researcher
September 2026$1,500$11,000
October 2026$2,400$23,000
Daily agent token spend per researcher at API list prices, September versus October 2026, in US dollars. Source: OpenAI, Research acceleration: The jagged frontier, November 18, 2026.1

And the uplift is jagged. Agents finished 46% of two-to-four-hour tasks unaided in October, up from 23% in March, but only 25% of inference and serving code 16.

Agents finish monitoring jobs, not the code that ships models

0%20%40%60%80%Monitoring training runs79%Answering technical questions72%Communicate69%Analyze58%Build43%Launching training runs35%Building evals and RL environments29%Inference and serving code25%
Data
Value
Monitoring training runs79%
Answering technical questions72%
Communicate69%
Analyze58%
Build43%
Launching training runs35%
Building evals and RL environments29%
Inference and serving code25%
Share of tasks agents completed with no human intervention, by task type, in percent. Source: OpenAI, Research acceleration: The jagged frontier, November 18, 2026.1

Holders pay for it, researchers carry the risk

OpenAI's holders pay the burn; this landed the week the lab raised $15 billion at a $1.25 trillion valuation, half what it first sought 3, while Anthropic$350B — Anthropic, private, latest valuation $350B priced its IPO at $83 and debuted on Nasdaq today at about $1.5 trillion fully diluted 5. Sellers of inference, from Nvidia down to the neoclouds, capture it. Researchers carry the other exposure: their own median answer reprices their colleagues' jobs.

OpenAI's counterpoint is fair: the headcount figure is a hypothetical preference, humans remain central, and Safety and Alignment workweeks grew 77% from January to September 1. Watch the next report's ratio. If agent spend keeps doubling against a wage bill it already matches, the question stops being whether agents accelerate research and becomes what the researchers are for.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading NVDA

Reader comments

0 comments

    Model

    Google's Gemini 4 Neon took the coding lead the day before Anthropic's $1.5 trillion debut

    Google's mid-tier model debuted at No. 1 on Artificial Analysis's coding-agent index at $6.86 a task against Claude Opus 5.5's $12.61, a day before Anthropic's Nasdaq debut. A $1.5 trillion valuation built on coding-led growth now meets a measured, cheaper rival, and the exposure sits with whoever paid $83.

    Model

    ElevenLabs, valued at $22 billion, puts its voice agents inside NICE, Genesys, Amazon Connect and Five9

    Salesforce and Five9 have put more than $3 billion into buying customer-service AI this month. ElevenLabs, valued at $22 billion after a September share sale, embedded its agents natively into Five9's platform and three rivals instead, and the evidence points to pricing power sitting with the channel, not the model maker.

    Model

    Anthropic's Claude and OpenAI's GPT-6.2 each top a 3D leaderboard during Anthropic's IPO week

    Anthropic, which is expected to price its Nasdaq IPO tonight, holds the crown on one human-vote 3D leaderboard and OpenAI holds the other. Spline, the vendor running the first board, pays its voters in its own AI credits and may hand their votes to research partners.

    Model

    HeyGen's Avatar 6 scored 84% on realism, but its 42.2% pick rate was below chance

    Avatar 6 launched November 17, 2026, with an 84% realism figure and first place in a blind arena; the score doubles a below-chance pick rate, and the arena capped 97% of clips at 35 seconds. That is the evidence base under a company claiming $200 million of annual revenue.

    Model

    MiniMax hit a record low, about 90% below its January peak, as a US rule gates its rented Nvidia compute

    Washington's new remote-access rule puts a US license between China's listed model makers and the offshore Nvidia capacity they rent. MiniMax, the purest listed bet on Chinese foundation models, holds the price tag.

    Model

    OpenAI cuts flagship pricing 20% to match Anthropic as heavy users call GPT-6.2 Astra unusable

    OpenAI cut its flagship's price to match Claude Fable 5.2's rate, and the users who run models hardest rejected it, 24 hours before Anthropic priced a $1.4 trillion IPO on the assumption the quality gap holds. The independent labs say the gap is real, for now.

    Subscribe to TheInference for free