OpenAI's own numbers show median internal use of its AI coding tool Codex jumped 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal between November 2025 and June 2026. The bottleneck, the data suggests, was never access. It was capability fit.
OpenAI gave every employee unlimited access to its own AI coding assistant, Codex. For most of 2025, almost nobody used it. Through August, the average OpenAI worker spent less than 10% of their tokens on the tool, according to the company's Economic Research team (OpenAI's internal Codex usage, reported by Latent Space).
Then, starting in November 2025, usage exploded. By June 2026, median Codex output tokens had grown 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal, OpenAI says in a new post on how agents are reshaping its internal workflows (OpenAI, "How agents are transforming work").
The pattern suggests something the AI rollout debate often misses. Unlimited access is necessary but nowhere near sufficient. What flipped the curve between August and November was not a procurement decision or a seat-expansion policy. OpenAI employees had all the tokens they could spend from day one. The shift tracks with Codex's own maturation as a tool. The moment it became useful in real workflows, the people who built it started using it like air.
That sequence matters for a specific reason. A common concern in late-2025 enterprise AI discourse was "tokenmaxxing," the worry that workers would burn unlimited AI tokens without producing proportional value. Industry commentary on X flagged the data as a stress test of that theory (commentary on X). OpenAI's own numbers do not support the runaway reading. The story inside the company is the opposite: nine months of under-use, then a sharp ramp. The bottleneck was never access.
The breakdown by department tells a second story about which workflows Codex actually fits. Research went 56x, a department whose day-to-day work involves reading, summarizing, and generating code. Customer Support went 32x, a function built around volume and pattern recognition. Engineering went 27x, lower than you might expect from the team that builds the model itself, though the multiplier applies to medians among active users, not the whole org. Legal went 13x, a function historically slower to adopt new tooling and whose documents are sensitive enough that safety rails likely cap some use cases.
For non-OpenAI teams, the diagnostic frame is concrete. If a rollout is stalled, do not blame budget or seat allocation first. Audit whether the tool actually fits the workflows employees are trying to run. OpenAI's own data says access was a non-issue from day one. What changed was the fit.
Two caveats belong on the record. The numbers are medians among active internal Codex users, not averages across the whole company, and they are self-reported by OpenAI Economic Research, not independently audited. The November 2025 baseline was low, which inflates the multipliers. There is no comparable external-customer or enterprise benchmark in the source bundle, so the analysis stays on internal adoption shape rather than claiming enterprise-wide parity.
The thing to watch next is whether the curve flattens by department, and whether the same adoption gap, months of under-use before a sharp ramp, shows up in other companies that gave employees unlimited AI access. The tokenmaxxing debate will not be settled by OpenAI's internal data alone. But inside the company that builds the model, the answer to "did unlimited tokens lead to runaway usage?" is, for nine months, no. Then, suddenly, yes.