OpenAI grants Codex customers a banked usage reset
OpenAI gave paid ChatGPT Work and Codex users a banked usage reset and said Codex has reached 20 million active users. The company is investigating reports that lower cache-hit rates are causing usage limits to drain faster.
TL;DR
- Paid Codex and ChatGPT Work users received a banked reset that they can hold and use later, according to thsottiaux's announcement.
- Delivery was set for 8 PM PST in thsottiaux's delivery update, and thsottiaux's confirmation later said the reset had landed.
- Codex reached 20 million active users during the week, thsottiaux wrote.
- A lower cache-hit rate for some users may account for faster-draining usage, according to thsottiaux's cache update.
OpenAI's Codex rate card says it moved most plans from per-message estimates to API token usage in April. Its current GPT-5.6 Sol row prices cached input at 12.5 credits per million tokens, compared with 125 credits for ordinary input.
Banked reset
OpenAI said paid Codex and ChatGPT Work users would be credited with a banked reset usable at their own leisure in thsottiaux's announcement. The company paired the credit with its 20 million active-user milestone for Codex.
In thsottiaux's delivery update, the company set an 8 PM PST delivery target for all paid users of the two products. It then said the reset was live in thsottiaux's confirmation.
Cache hits
In thsottiaux's first update, OpenAI said it was not seeing anything abnormal while investigating complaints about usage draining unusually fast. A participant in an OpenAI Developer Community post described a workload that previously ran across multiple sessions reaching its weekly limit after one project.
The investigation subsequently found that some users had worse cache-hit rates than in the preceding stable period, thsottiaux's cache update said. The post framed that degradation as a possible explanation for the faster burn.
The rate card meters input, cached input, and output separately. On GPT-5.6 Sol, an ordinary input token costs ten times the cached-input rate, so a cache miss can sharply raise the input-side debit even though output tokens continue to count separately.
Shared agentic pool
Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draw from the same agentic usage and credit pool when those features are available on a plan, the rate card says. It estimates a typical GPT-5.6 Sol Codex task at 5 to 40 credits, while noting that model choice, token mix, concurrent instances, automations, and fast mode change the total.
The same page says Codex's Settings > Usage panel can show limits and, depending on plan and workspace role, remaining credit, purchases, and auto-reload controls. OpenAI's desktop usage guide adds that eligible Enterprise and Edu users can see Work and Codex history, plus locally available high-usage chats and per-chat details where available.
A bcherny reply named extreme parallelism, runaway loops, and inefficient skills or plugins as common sources of high consumption, and said /usage returns a detailed breakdown.