OpenAI reports internal Codex token usage surged 56x in Research since November 2025
Median output tokens for Codex grew sharply across departments, with Research seeing the largest increase. The data reflects internal agentic workflows becoming integral to operations.
1 source · cross-referenced
- OpenAI's internal data shows median Codex output tokens grew 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since November 2025.
- The surge indicates agentic workflows are increasingly embedded in cross-functional tasks at OpenAI.
- Research teams led the adoption, with median usage far outpacing other departments.
OpenAI’s internal reporting indicates that median output tokens for its Codex system have grown dramatically across multiple departments since November 2025. According to the data, Research saw the largest increase, with median tokens rising 56-fold by June 2026. Customer Support followed with a 32x increase, Engineering with 27x, and Legal with 13x.
The figures reflect a shift in how OpenAI employees are using Codex, with the tool becoming integral to longer-running, cross-functional tasks rather than isolated coding assistance. OpenAI Economic Research noted that, through August 2025, the average worker spent less than 10% of their tokens on Codex, indicating a rapid acceleration in adoption over the subsequent months.
The internal data underscores the growing role of agentic workflows within the organization. Commentary from OpenAI staff and external observers highlighted patterns such as the use of skills and concurrent agents, suggesting that internal tooling is evolving to support persistent, reviewable, and tool-integrated processes.
While OpenAI employees have had unlimited access to Codex, the sharp rise in usage signals a transition from underutilization to deep integration. The trend provides a baseline for evaluating concerns about "tokenmaxxing" and offers insight into how agentic systems are being operationalized in practice.
- Aug 6, 2026 · arXiv cs.AI
Paper proposes self-verifying agent instrument to isolate commitment and binding drift in long-horizon agents
Trust79 - Aug 5, 2026 · Latent Space — swyx
Cursor open-sources Mixture-of-Kittens megakernel for MoE training
Trust72 - Aug 4, 2026 · Hugging Face
Liquid AI releases LFM2.5-2.6B, a 2.6B-parameter on-device agentic model with tool-calling and multi-step workflows
Trust79