Engineering workflows are shifting as internal software tooling begins to consume more theoretical capital than the base salaries of the specialists operating it. According to OpenAI's second report on internal research acceleration, internal agent usage has expanded rapidly as the organization documents indicators of recursive self-improvement, where AI systems accelerate future development cycles. The resulting compute overhead creates a notable shift in development spending.
At regular API list prices, the median researcher at OpenAI consumed $2,400 worth of tokens per day in October, marking a 62 percent increase from the previous month. Among the top 10 percent of users, that theoretical value reached $23,000 per day. In September, the theoretical API equivalent of agent runtime was twice as high as researchers' regular salary costs, even though these calculations exclude equity packages that make up 70 to 80 percent of total employee compensation. This trajectory raises the question of whether autonomous coding agents can assume complex engineering burdens or whether they primarily multiply routine code volume.
Workday Multipliers and Parallel Compute
Operational capacity inside OpenAI has scaled aggressively around these automated systems. Converted into eight-hour workdays, internal agents run the equivalent of 5.7 workdays per researcher per calendar day, expanding to about eight workdays within pre- and post-training teams. During peak usage periods, more than two-thirds of tracked researchers deployed at least four agents in parallel.
Internal agents run the equivalent of 5.7 workdays per researcher per calendar day, expanding to about eight workdays within pre- and post-training teams.
This continuous execution alters staffing preferences across technical groups. In an internal, non-representative survey of 101 employees, researchers reported at the median that they would rather give up 40 percent of their team's staff or 30 percent of compute capacity than spend three months without their current coding agents. In a separate quick poll of 28 employees, 27 preferred the latest internal model over a mid-career human engineer.
Engineering Bottlenecks and Supervised Runtimes
Despite heavy reliance on automated tooling, performance varies sharply depending on task architecture. Between March and October, the estimated success rate of autonomous agents operating without human intervention on two- to four-hour tasks increased from 23 to 46 percent, while four- to eight-hour tasks rose from 13 to 35 percent. Tracking under an Epoch AI taxonomy shows agents handle structured routines effectively, completing training-run monitoring 79 percent of the time and answering technical questions independently 72 percent of the time.
Breakdowns occur when systems encounter core system architecture. Agent success rates drop to 29 percent for building evaluation and reinforcement-learning environments, and to 25 percent for writing code for inference and serving systems.
The models lack any "taste" for good research ideas and for code and data quality, according to researchers interviewed in the report.
Human direction remains central to core architecture. The key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra originated exclusively from humans in post-training and safety, with agent contributions limited to individual researcher dialogues with Codex during GPT-6.2 Astra pretraining.
An objection to these cost figures is that they rely on standard API list prices rather than OpenAI's actual marginal compute costs. Measured directly against OpenAI's total research compute infrastructure, internal agent use represented just 2 percent of total capacity. Industry tracking outside the company indicates broader acceleration trends, with evaluation organization METR estimating after an eleven-week evaluation that AI tools nearly doubled development pace at Anthropic. OpenAI had previously declared in early September that it met its internal goal of an automated AI research intern, though Chief Scientist Jakub Pachocki subsequently warned of losing control after an audit revealed multiple gaps in the oversight system.
While external pricing suggests median researcher token usage of $2,400 daily, internal agent execution consumes only 2 percent of OpenAI's total research compute.
What’s behind it
The operational reality points in the opposite direction: while engineers readily give up live colleagues for digital helpers, internal API-rated compute expenses have climbed to double the baseline researcher salaries. Even though the cost per million tokens fell by 97% by mid-July, corporate budgets saw no relief as teams simply started running up to eight shifts worth of automated runs every single day. Yet with success rates bottoming out near 25% on critical inference code, massive compute consumption still hits a hard ceiling whenever agents face core architecture. Autonomous tools effectively replaced routine development shifts at staggering internal prices, but researchers still have to step in whenever the code touches foundational infrastructure.
Our stories: OpenAI GPT-5 Strategy: Why ROI Trumps Token Costs (July 15).