OpenAI 研究员人均每天消耗 2400 美元 API 调用,GPT-6.2 关键洞察首次由 Codex 参与
OpenAI 内部数据首次把 agent 用量和产出挂钩:每天 2400 美元换来实验多 68%,但能力提升没跟上。最弱的是写服务代码,成功率只有 25%,这道短板卡住了整体加速。做科研自动化或者算 agent ROI 的团队都该看看这组数字。
OpenAI 内部数据显示,普通研究员每天在 AI agent 上的 API 开销达 2400 美元,最重度使用的 10% 研究员日均消耗超过 23000 美元。每名研究员的 agent 工作时长相当于 5.7 个人按 8 小时轮班,较此前增长 11 倍。
产出增长远低于投入:2025 年以来实验数量增加 68%,能力提升仅略高于行业整体趋势。值得注意的是,针对 GPT-6.2 Astra 的预训练研究员表示,Codex 帮助找到了关键洞察,这是 AI 首次被报告在“研究品味”层面发挥作用——而 GPT-6 Astra 的核心想法仍全部来自人类。
任务分层差异明显:agent 在监控训练任务上成功率达 79%,但写推理与推理服务代码的成功率只有 25%。报告指出,这些最弱环节可能成为整体科研加速的上限,无论 agent 使用量如何增长。
The typical OpenAI researcher now burns through $2,400 a day on AI agents at API prices, and the heaviest 10% run past $23,000 a day.
> Every key idea researchers named for GPT-6 Astra came from people, but for GPT-6.2 Astra, pretraining researchers said Codex helped find key insights, AI’s first reported foothold in research taste.
> Each OpenAI researcher has AI agents running as many hours as 5.7 people on 8-hour shifts every day, up 11-fold. Yet output rose far less: 68% more experiments than in 2025, and capability gains just above the industry trend.
> Agents succeed 79% of the time monitoring training runs but only 25% writing inference and serving code, and these weakest tasks could cap overall research speedup however much agent use grows.
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include OpenAI, GPT-6, Codex, AI agent.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).