OpenAI just gave us the first real price tag for living in the agent economy, and it's higher than most startups' entire cloud bill.
The Summary
- OpenAI researchers now burn $600/day median in AI tokens for coding agents, with power users hitting $7,000/day — up from $162/day in July
- Internal tech support posts dropped 50% since January as agents handle troubleshooting, showing AI replacing coordination overhead before it replaces jobs
- OpenAI claims they've hit the "automated research intern" milestone and are targeting fully automated AI researchers by March 2028
The Signal
The numbers tell you where this is headed. When your median researcher is spending $600 a day on tokens and nobody's blinking, you're not optimizing workflows anymore. You're building a different cost structure for knowledge work entirely.
Think about what $600/day means. That's $156,000 per researcher per year, just in token spend. For context, that's roughly what OpenAI pays a junior engineer in total compensation. The company is effectively running two payrolls: one for humans, one for the compute those humans direct.
"The median researcher at OpenAI is now managing a six-figure annual AI budget on top of their salary."
The velocity tells you more than the absolute numbers. July to August saw a 270% jump in median daily token spend. That's not incremental adoption. That's a phase change in how work gets done. The 90th percentile users quintupling their spend in the same window suggests the top researchers have figured something out that's now cascading down.
OpenAI buried the real story in the support channel data. When internal tech support posts drop 50% because agents are fielding the questions, you're watching coordination costs collapse. That's not automation replacing a job. That's automation eliminating the friction that made the job necessary. The agent doesn't need to be better than the support engineer. It just needs to be faster than waiting for one.
Key dynamics at play:
- Token spend is becoming the new cloud infrastructure line item, except it scales with cognitive load instead of data storage
- Power users spending 10x the median suggests steep learning curves on prompt engineering and agent orchestration
- The support channel metric shows agents eating coordination work first, not creative work
The March 2028 timeline for "fully automated AI researchers" is classic Altman goal-setting. Ambitious enough to move the org, specific enough to measure. But the real milestone already happened. They hit "automated research intern" this year. That means agents are now doing bounded tasks end-to-end: write the test, run the experiment, document the results. Human researchers are becoming research directors.
This is what the agent economy actually looks like at the frontier. It's not AI replacing workers. It's AI becoming the variable cost of getting work done, with humans as the fixed cost that directs it. When a researcher can spend $7,000 in a day and ship code faster than they could alone, the unit economics of R&D just changed. You're not hiring for output anymore. You're hiring for judgment about what to build and how to orchestrate agents to build it.
The Implication
If you're a developer, the play is obvious: learn to direct agents at scale, not just write code. The researchers spending $7,000/day aren't coding more. They're architecting systems of agents that code for them. That's the skill that commands a premium now.
For companies, this is your new budgeting reality. Token spend will become a per-employee line item, not a team resource. Plan for researchers, analysts, and engineers to carry five-figure to six-figure annual AI budgets. The companies that figure out ROI measurement on agent spend first will out-execute everyone else.
Watch what happens when OpenAI's March 2028 milestone hits. If they actually build fully automated AI researchers, the entire innovation stack reorganizes around humans who can pose the right problems and evaluate novel solutions. That's a different job than researcher. We don't have a name for it yet.