OpenAI's median researcher used about $2,400 a day of agentic coding at API list prices in October, and agents finished only a quarter to a third of the hard engineering tasks without human help. For anyone setting a 2027 agent budget per engineer, it is one of the most specific per-person spend figures a frontier lab has published, and the success rates matter more than the spend: the money buys the most on monitoring and analysis work, the least on production code, and every dollar of output lands on someone's review queue.
The numbers come from OpenAI's "Research acceleration: The jagged frontier" report, published November 18, 2026. It is OpenAI describing its own staff, with OpenAI's incentives, and it says so. It is still one of the most detailed public looks at what heavy agent use costs and returns.
What OpenAI Measured in October
OpenAI reports a median of about $2,400 per researcher per day at API list prices, up 62% from $1,500 in September, with the 90th percentile at about $23,000 a day, up from $11,000, according to the November report. Researchers there have no agentic coding quotas.
The growth curve is steep. OpenAI's September report put the median at "more than $600 per day of inference at API prices" by mid-August, and the 90th percentile at more than $7,000. Part of the jump since then is an artifact: the November report's footnote says a temporary harness efficiency change deployed internally on August 21 "appears to have severely depressed" documented spend, and it was reverted gradually from September 10. Some of the August-to-October rise is that change unwinding; OpenAI says the trend also reflects more activity, shifts to more expensive models and changes to model behavior and harnesses.
Two other figures describe the volume behind the spend. OpenAI defines an "agent workday" as eight hours of agent use, and the average per researcher per day rose from 0.5 to 5.7. By mid-October, more than two-thirds of the researchers it counted averaged a daily peak of at least four concurrent agents, and more than 10% averaged at least sixteen. The same report says the API-equivalent value of that use was, as of September, about twice researchers' non-equity compensation.
Why $2,400 a Day Overstates Your Bill
The $2,400 figure is a ceiling for planning, and three details in the report keep it from being a forecast for your team.
First, it is priced at list. OpenAI is the vendor, and the report values its usage at API list prices, not at what the compute costs OpenAI; internal deployment is "only 2% of total research compute," per the report. Your rate will be a negotiated or plan rate. On seat plans the math works differently again: OpenAI's Codex pricing docs say "ChatGPT Work and Codex share pricing, credits, and usage limits," so a seat buys a capped allowance with paid credits on top.
Second, the pay comparison excludes equity, which the report says "generally accounts for 70 to 80% of target compensation" at OpenAI. Agent spend at twice cash pay is a much smaller share of total pay. For a reference point outside a frontier lab, the US Bureau of Labor Statistics puts the median software developer wage at $135,980 (May 2025). If $2,400 ran every working day (about 250 a year), it would come to roughly $600,000 at list price, more than four times that wage. No enterprise CFO will sign that, and nothing in the report suggests you should.
Third, these are researchers training frontier models, with unlimited access and every incentive to push agents hard. Your engineers maintain a billing system. The steel-man case for the report is that it shows where usage goes when cost stops being a constraint, which is a useful upper bound. Use it as that.
Which Tasks Agents Finish Without Help
OpenAI's success rates split cleanly by task type, and the split tells you where to send agent budget first. Agents completed these tasks with zero human intervention at the following estimated rates, per the report. The 35% launch figure comes from the report's summary, which gives no date window; the others cover August 1 to October 18, 2026.
| Task | Unassisted success |
|---|---|
| Monitoring runs for issues | 79% |
| Answering technical questions | 72% |
| Launching training runs | 35% |
| Building eval, RL and agent environments | 29% |
| Writing code for inference and serving stacks | 25% |
By broad category, using the task taxonomy from Epoch AI's "O*NET for AI R&D", Communicate tasks succeeded 69% of the time, Analyze 58% and Build 43%. By estimated human duration, success on 4-8 hour tasks rose from 13% to 35% between March and October. OpenAI flags that researchers may have changed which tasks they delegated, so the rise "may reflect both capability improvements and researchers' changing choices."
Map that onto an enterprise engineering org and the order is clear. Watching dashboards, triaging alerts, summarizing logs and answering "where does this config live" questions are the monitoring and question-answering rows. Production serving code, deployment launches and test harnesses are the bottom three. OpenAI's own report says lower-performing categories "could become bottlenecks to high overall uplift," which is the point: an agent that clears 79% of triage but 25% of serving changes speeds up the part of the job that was rarely the constraint.
OpenAI's operations team shows the same pattern in production. In the Performance Factory write-up, a regression workflow running Codex agents tracked more than 1,700 investigations and root-caused over 600 in just over three months. In a manual review of 100 root-cause suggestions, 70% were graded accurate. Fixes still go out as pull requests: "An engineer reviews the PR and decides whether to merge it." We covered that system in OpenAI Uses Codex for Diagnosis, Engineers for ChatGPT Fixes.
Review Capacity Becomes the Constraint
The scarce resource in both labs' data is the human reading the output. OpenAI's interviews say "Workflows require close human involvement to avoid low-quality outputs and wasted cycles," and researchers described agents that rabbit-holed, wrote "AI slop" PRs and made invalid analysis assumptions, per the report. One researcher summarized the working pattern as "Go make one. I'll review it and make edits. Then go do the rest."
Anthropic's numbers say the same thing from the other side. Its October paper on how employment is changing at Anthropic found Claude sessions are "around 80% idly waiting for human input," and that reading all interactive output would take non-technical groups over half the work week. Code commits per engineer rose to 5x their level of six months earlier. A separate Anthropic measurement page reports about 30,000 agents running at any time on its main internal platform as of August 2026, with roughly 100,000 transcripts flagged per week and about 50 a week escalated to humans: the lab built automated monitoring precisely because people cannot read everything.
Independent assessment is thinner. METR's review of AI R&D acceleration at Anthropic estimated capability gains on well-specified tasks at around twice the earlier rate, and noted that Anthropic redacted a majority of the evidence METR found most informative. Self-reported speedups in Anthropic's Slack polls reached 12x in October, which METR called "only very loose evidence." Treat any vendor's internal uplift number the same way.
For your org, this changes the budgeting unit. If concurrency per engineer goes from one agent to four or sixteen, PR volume rises and review hours do not. The Harvard and Jellyfish study we covered found the same review bottleneck in commercial engineering teams.
How to Size a 2027 Agent Budget Per Engineer
Size agent spend by task type and cap it at what your reviewers can absorb, then fund review capacity as a line item next to tokens. OpenAI's data supports routing budget toward high-success work first and treating low-success work as a pilot.
This Week:
- Pull last month's agent spend per engineer from your provider's admin console and get the median and 90th percentile on paper. Compare both to OpenAI's $2,400 and $23,000 list-price figures, so you know whether you are under-using agents or just under-reporting.
- Ask your platform lead to tag a sample of 50 recent agent sessions by the five task rows in the table above. That gives you your own success-by-type split instead of OpenAI's.
This Month:
- Route monitoring, alert triage and codebase Q&A to agents by default; those are the 72-79% rows. Keep serving-path code, deploy launches and eval harness work behind a named human reviewer.
- Measure reviewer hours per merged agent PR. If that number rises as agent count rises, more token budget buys a longer queue. A tool from our AI coding ROI measurement comparison can pull this from your Git history.
- Model your bill on your actual plan. Seat plans for Codex and GitHub Copilot cap usage and charge for overage differently from raw API use; our cost per developer breakdown has the numbers.
Before Budget Lock:
- Write the 2027 line as two numbers: agent spend per engineer and review hours per engineer. A budget that only funds tokens will hit the review constraint by spring.
- Set the per-engineer ceiling as a multiple of loaded salary that finance agrees to now. OpenAI's ratio of about 2x cash pay is a lab with no quotas; pick yours on purpose and revisit it quarterly, since OpenAI's own median rose 62% from September to October, partly as the harness change unwound.
The Bottom Line
The report is OpenAI grading itself, priced at API list rates rather than OpenAI's own cost, for a workforce unlike yours. What transfers is the shape of the curve. Uncapped, the median rose from about $600 to $2,400 a day between mid-August and October (part of that a measurement artifact), and agents did well on watching and explaining while doing badly on shipping production code.
Cloud went the same way a decade ago: once servers were cheap to provision, the constraint moved to whoever had to watch the bill. With agents it moves to whoever reads the pull requests. Budget the reviewer before you budget the next agent.
Continue Reading
- OpenAI Uses Codex for Diagnosis, Engineers for ChatGPT Fixes
- AI Coding Agents Lift Pull Requests 35.7%, Harvard Study Finds
- Your Copilot Bill Can Reach $137 a Month at 10 Daily Tasks
- Choose Swarmia Over LinearB for AI Coding ROI PR Comparisons
- ChatGPT and Codex Will Share OpenAI's Usage Allowance in December
- AutoBuddy Finds Agents Spot Owed Work but Fail During Execution
