OpenAI Says Its Researchers Now Burn Up to $7,000 a Day on AI Tokens

OpenAI reported on September 6 that its median researcher now spends over $600 per day on AI tokens, up from $162 in July, with the top 10% spending more than $7,000 daily. The company claims its research organization logs 3.1 'agent-workdays' for every eight-hour human workday, a milestone CEO Sam Altman had targeted for September 2026. However, no outside lab or auditor has independently verified these figures, and over half of successful agent tasks lasting four to eight hours still required human intervention.

OpenAI's own researchers are spending more on AI tokens than most people spend on rent, and the company is calling it a milestone rather than a warning sign. On September 6, OpenAI published a report titled "Research acceleration: the view inside OpenAI," and the numbers inside are startling even by Silicon Valley standards. The median researcher at the company now spends more than $600 a day running coding agents against OpenAI's own models, up from just $162 a day in July. The heaviest 10% of users blow through more than $7,000 a day. That's not a typo. That's one researcher's daily token bill rivaling a decent monthly salary. OpenAI frames this spending spike as proof of something bigger: its research organization now logs 3.1 "agent-workdays" of coding and experimentation effort for every single eight-hour day a human researcher puts in. The company says it crossed that threshold only since June. In plain terms, OpenAI is telling the world that AI agents inside its own building are now doing more than three times the raw work volume of the humans who built them. Sam Altman set this exact target last October. During an October 2025 livestream, he said it was "plausible" that by September 2026, OpenAI would have an intern-level AI research assistant. A full AI researcher, he said, was still years off - March 2028. OpenAI's new report claims it hit that first mark on schedule. The milestone: a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days to finish. The report also notes something else. A growing share of researchers now run four or more agents at once, treating them less like autocomplete and more like a team of junior staff working in parallel. But there's a catch buried in the same document: over half of the successful agent tasks that ran four to eight hours still needed at least one human intervention to land. The intern still needs a manager checking in. OpenAI Chief Scientist Jakub Pachocki Warns No AI Lab Is Ready to Scale Safely https://startupfortune.com/openai-chief-scientist-jakub-pachocki-warns-no-ai-lab-is-ready-to-scale-safely/ OpenAI chief scientist Jakub Pachocki published an essay saying no AI lab, including OpenAI, has solved alignment well enough to keep scaling at full speed, just three days after OpenAI shipped GPT-6 Astra, its first model to cross a Critical cybersecurity threshold. Sam Altman called the essay 'an important post' even as Astra keeps rolling out... - how to safely scale autonomous AI systems https://startupfortune.com/openai-chief-scientist-jakub-pachocki-warns-no-ai-lab-is-ready-to-scale-safely/ - AI lab safety concerns for frontier models https://startupfortune.com/openai-chief-scientist-jakub-pachocki-warns-no-ai-lab-is-ready-to-scale-safely/ Here's the thing. Nobody outside OpenAI verified any of this. OpenAI wrote the definition of "automated research intern," OpenAI ran the measurements, and OpenAI published the grade. No outside lab, no academic auditor, no independent benchmark confirmed the 3.1 ratio or the productivity gains behind it. Treat the claim the way you'd treat any report where the subject and the grader are the same company. That skepticism doesn't erase the spending data, though. The token bills are real, denominated at standard OpenAI API retail rates, and they tell founders and investors something concrete regardless of how you feel about the productivity claims layered on top. What this means for AI coding tool spend elsewhere If OpenAI's own researchers, who get internal access and presumably favorable rates, are running up bills like this, founders paying full price for Cursor, Devin, or GitHub Copilot Workspace should expect their own agent spend to climb the same curve. It's already happening. The jump from $162 to $600 a day happened in roughly one month, right around when internal staff reportedly got early access to a more capable model. That's the pattern to watch: every time a materially stronger model ships, token consumption per researcher doesn't rise gradually, it jumps. For a startup budgeting AI tooling costs, that's the real lesson buried under the Also read: UK Minister Admits Palantir Mistrust Is Driving NHS Patients to Opt Out https://startupfortune.com/uk-minister-admits-palantir-mistrust-is-driving-nhs-patients-to-opt-out/ • Alibaba's Qwen3.8-Max Model Overtakes Claude Opus 5 on Coding Leaderboard https://startupfortune.com/alibabas-qwen38-max-model-overtakes-claude-opus-5-on-coding-leaderboard/ • China's Grip on Indium Phosphide Is Becoming a Real Problem for AI Chipmakers https://startupfortune.com/chinas-grip-on-indium-phosphide-is-becoming-a-real-problem-for-ai-chipmakers/ Join the discussion Open in the community → /community/ Almost there. Sign in and your reply posts straight away.