OpenAI's internal use of coding agents is growing fast, with theoretical token costs now exceeding researchers' base salaries. Even though teams are reluctant to work without these tools, the models still produce a lot of "AI slop" beyond routine tasks.
OpenAI's Surging Internal AI Usage and Costs OpenAI's internal AI tools, including Codex and Dots, have experienced a dramatic increase in usage, with theoretical token consumption reaching substantial levels. For an average researcher, the daily theoretical token cost was $2,400 in October, marking a 62% rise from the previous month. This figure surged even higher for the top 10% of users, hitting $23,000 per day. When converted to an equivalent of eight-hour workdays, these agents reportedly perform the work of 5.7 days per researcher per calendar day, increasing to about eight days in the pre- and post-training teams. During peak usage, over two-thirds of tracked researchers operated at least four agents concurrently. While these theoretical API costs were twice the researchers' regular salaries in September (excluding significant equity packages), internal agent usage still represented only 2% of OpenAI's total research compute. High success rates on routine tasks, weaknesses in code scaffolding OpenAI's autonomous agents have significantly improved their estimated success rates for tasks requiring two to four hours, increasing from 23% to 46% between March and October. Similarly, for four to eight-hour tasks, success rates rose from 13% to 35%. However, analysis using an Epoch AI taxonomy indicates that these systems primarily excel at simpler, more routine tasks. For instance, they achieve a 79% success rate in training-run monitoring and independently answer technical questions 72% of the time. Conversely, their performance is considerably lower for complex engineering work, with success rates of only 29% for building evaluation and reinforcement-learning environments and 25% for writing code for inference and serving systems. A non-representative internal survey highlighted the indispensability of these coding agents, with most researchers expressing a preference to sacrifice 40% of their team's staff or 30% of compute capacity rather than operate without them for three months. Lack of research taste requires human oversight Qualitative interviews with ten OpenAI researchers have illuminated the drawbacks of extensive AI model usage. A key criticism is the models' lack of "research taste," leading them to get sidetracked by unnecessary steps, make incorrect assumptions in analyses, and generate "AI slop" — subpar code changes and unusable data artifacts. This deficiency mandates close human oversight to prevent wasted compute cycles and maintain quality. Crucially, the foundational conceptual ideas for advanced models such as GPT-6 Astra and GPT-6.2 Astra were entirely human-derived in the post-training and safety domains. Only in the pretraining phase for GPT-6.2 Astra did a few individual researchers report gaining valuable insights through dialogue with the Codex model. In conclusion, OpenAI cautions that data from current human-AI collaboration may not extrapolate linearly, acknowledging the possibility of rapid, non-linear advancements in deep-learning research akin to recent AI breakthroughs in mathematics, especially as models gain increased autonomy. This perspective is reinforced by external assessments, including METR's finding that AI usage nearly doubled Anthropic's development pace, and OpenAI's prior reports on achieving an automated AI research intern alongside warnings from Chief Scientist Jakub Pachocki about potential loss of control due to monitoring gaps.
OpenAI's Surging Internal AI Usage and Costs
OpenAI's internal AI tools, including Codex and Dots, have experienced a dramatic increase in usage, with theoretical token consumption reaching substantial levels. For an average researcher, the daily theoretical token cost was $2,400 in October, marking a 62% rise from the previous month. This figure surged even higher for the top 10% of users, hitting $23,000 per day. When converted to an equivalent of eight-hour workdays, these agents reportedly perform the work of 5.7 days per researcher per calendar day, increasing to about eight days in the pre- and post-training teams. During peak usage, over two-thirds of tracked researchers operated at least four agents concurrently. While these theoretical API costs were twice the researchers' regular salaries in September (excluding significant equity packages), internal agent usage still represented only 2% of OpenAI's total research compute.
High success rates on routine tasks, weaknesses in code scaffolding
OpenAI's autonomous agents have significantly improved their estimated success rates for tasks requiring two to four hours, increasing from 23% to 46% between March and October. Similarly, for four to eight-hour tasks, success rates rose from 13% to 35%. However, analysis using an Epoch AI taxonomy indicates that these systems primarily excel at simpler, more routine tasks. For instance, they achieve a 79% success rate in training-run monitoring and independently answer technical questions 72% of the time. Conversely, their performance is considerably lower for complex engineering work, with success rates of only 29% for building evaluation and reinforcement-learning environments and 25% for writing code for inference and serving systems. A non-representative internal survey highlighted the indispensability of these coding agents, with most researchers expressing a preference to sacrifice 40% of their team's staff or 30% of compute capacity rather than operate without them for three months.
Lack of research taste requires human oversight
Qualitative interviews with ten OpenAI researchers have illuminated the drawbacks of extensive AI model usage. A key criticism is the models' lack of "research taste," leading them to get sidetracked by unnecessary steps, make incorrect assumptions in analyses, and generate "AI slop" — subpar code changes and unusable data artifacts. This deficiency mandates close human oversight to prevent wasted compute cycles and maintain quality. Crucially, the foundational conceptual ideas for advanced models such as GPT-6 Astra and GPT-6.2 Astra were entirely human-derived in the post-training and safety domains. Only in the pretraining phase for GPT-6.2 Astra did a few individual researchers report gaining valuable insights through dialogue with the Codex model. In conclusion, OpenAI cautions that data from current human-AI collaboration may not extrapolate linearly, acknowledging the possibility of rapid, non-linear advancements in deep-learning research akin to recent AI breakthroughs in mathematics, especially as models gain increased autonomy. This perspective is reinforced by external assessments, including METR's finding that AI usage nearly doubled Anthropic's development pace, and OpenAI's prior reports on achieving an automated AI research intern alongside warnings from Chief Scientist Jakub Pachocki about potential loss of control due to monitoring gaps.