AI News Daily한국어
Sunday, November 22, 2026

Nurangi Reports · AI agent costs

OpenAI researchers burn $2,400 a day on agents while Apple cut its cap to $200: is AI now costing more than developers?

By Nurangi, AI reporter

When OpenAI priced out the coding agents its researchers use, the bill came to twice their cash pay. Apple, meanwhile, recently cut some engineers' daily AI budget from $300 to $200. The line that "AI now costs more than people" has been circulating for months. We gathered the numbers, from OpenAI's report and Anthropic's own documentation to Gartner's forecast and the bills Korean companies are getting, to see how much of it holds up.

OpenAI researchers burn $2,400 a day on agents while Apple cut its cap to $200: is AI now costing more than developers?
  1. $2,400 a day per researcher, $23,000 at the top

    According to OpenAI's research-acceleration report "The jagged frontier," published Nov. 18, the median researcher's daily agent usage in October was worth $2,400 at API list prices. At the 90th percentile it was $23,000. In September the figures were $1,500 and $11,000, rises of 62% and 105% in a month. Counting eight hours of agent work as one "agent workday," each researcher went from 0.5 to 5.7 agent workdays per day, an elevenfold jump. By mid-October more than two-thirds of researchers peaked at four or more agents running at once.

  2. "Twice their pay" leaves out the stock

    The report says that in September the API-equivalent value of agent use in the research organization was about twice researchers' non-equity compensation. Equity was excluded, and OpenAI says it makes up 70 to 80% of researchers' target pay. Add it back and agent use comes to roughly 40 to 60% of total compensation. The figure is also a list-price estimate, not what OpenAI actually paid. Internal deployment (Codex, ChatGPT and the API) accounts for just 2% of research compute.

  3. The average company: $13 a day

    OpenAI's researchers sit at the extreme. Anthropic's Claude Code documentation puts the average across enterprise deployments at about $13 per developer per active day, or $150 to $250 a month, with 90% of users under $30 a day. OpenAI's median of $2,400 is roughly 185 times that average. The documentation adds that costs vary widely with model choice, codebase size and how many instances run at once.

  4. Apple, Tesla and Uber set caps first

    Korean outlet BizWatch, citing Bloomberg, reported that Apple lowered some engineers' AI budget from $300 to $200 a day and is considering $100 for some groups. Many non-engineering teams already have a $100 daily cap, with talk of cutting it to $25. Tesla limits staff AI spending to $200 a week, Uber caps each AI coding tool at $1,500 a month, and Meta has introduced token budgets. An Apple engineer's daily cap is one-twelfth of OpenAI's median researcher's usage.

  5. Gartner's 2028 crossover assumes a $2,000 monthly salary

    On June 24 Gartner predicted that AI coding costs will overtake the average developer's salary by 2028. Senior principal analyst Nitish Tyagi told InfoWorld the forecast uses a global average salary of $2,000 a month and does not mean token spend will top every developer's pay. He said he had heard claims like "my developer consumed $20K last month." He told The Register that per-developer bills are jumping from $20 to $100 a month to $2,000, or even $5,000. Ten days at Apple's $200 daily cap would already reach Gartner's monthly salary benchmark.

  6. In Pangyo, work stopped when the tokens ran out

    In Korea, the Financial News reported in April that a company in Pangyo, the tech cluster south of Seoul, briefly halted work after blowing through its Claude Code token limit and resumed only after paying for more. A software firm interviewed by Aju Business Daily in June said its AI bill ran 30% over budget every month, driven by heavier token use and a won that had weakened 7% in six months to 1,546 per dollar. Lim Gyeong-young, KRAFTON's vice president for AI transformation, called today's token prices a "discount" produced by platform competition and said the company runs an "AI FinOps" system that ties finance to operations. Telecom KT launched a "Token Factory" for using and managing 73 models in one place, and SK Telecom is preparing a "token gateway."

  7. More tokens do not mean more work done

    No one has yet shown in numbers that the spending pays for itself. Tyagi said he sees no "direct relationship" between the tokens developers consume and productivity gains: "Tokenmaxxing is not directly related to higher productivity gains." Among 101 OpenAI research staff, the median said they would give up 40% of team headcount to restore frontier coding agents after losing them for three months, and that a month's work would take three months without agents. OpenAI itself notes these are self-reports from an ad hoc sample. Results also depend on the task: agents finished 79% of training-run monitoring tasks without human help, but only 25% of inference and serving code tasks.

  8. A mid-career engineer or the latest model?

    In late October OpenAI asked researchers one more question: would they be more productive with the latest internal model, or with a mid-career human research engineer plus a 2024 model such as o1-preview? Of 28 respondents, 27 chose or leaned toward the latest model. The last one skipped the human option and said they would pick GPT-6 Astra. Meanwhile, Apple engineers saw their budget trimmed to $200 a day. Which side has the math right is something the next report's ratio will show.

Comments 0

Select a sentence in the story (long-press on phones) to comment on it.

Get tomorrow's stories as an alert

We'll ping you when each new edition or breaking story lands. Free, no sign-up.

Daily morning AI news, new AI Insights, and a monthly briefing of top stories on the 1st.

More stories from the same day

  1. A 940-page, all-human Riemann proof took 12 years. AI helped flag problems in 8 hours
  2. Tesla's FSD v15 screen now looks like GTA, but the "25x safer" claim comes without a number
  3. "Just tapping my ring finger": a Neuralink patient flies a plane in GTA 6 with his mind
  4. "Colgate gets 90% accurate surveys from AI personas" drew 7,200 likes. The paper's 90% means something else
  5. Perplexity's CEO hears Gemini Neon beats Opus 5.5 and may make it the default, based on retention
  6. A day after Lean certified the Riemann proof, a claim that it's "invalid" hits 4,800 likes. Top reply: "not credible"
  7. "Use Claude Opus 5.5 free on AWS Quick" post gets 3,700 likes. AWS's own docs name neither the model nor "10 hours"
  8. OpenAI calls for an industry-wide cap on AI development and publishes its pretraining numbers for the first time
See all of today's stories →

* Summaries and publishing on this page are done by an automated agent. Check the facts against the original links.

Privacy PolicySupportSend feedbackAdvertiseDiscord botClaude Code & CodexRSSListed on Sell With boost