How many Codex credits does a task cost
OpenAI states GPT-5.6 usage averages 5 to 40 credits per message. OpenAI states GPT-5.6 usage averages 5 to 40 credits per message; fast mode increases consumption.
Credits by workload size, GPT-5.6 Terra
Defined workload baskets, not measured tasks: a short chat turn, a typical daily-coding request, and one step of a longer agent session, each with a different input size and cached share.
| Workload | Credits | Cost |
|---|---|---|
| Short chat turn | 0.5 | $0.02 |
| Daily coding request | 1.0 | $0.04 |
| Agent session step | 3.3 | $0.13 |
Numbers above are for GPT-5.6 Terra; a cheaper or pricier model shifts every row by the same ratio. See all models for the full comparison.
Why the range is wide
A message's credit cost is a function of tokens, not a flat per-interaction fee: a one-line question and a message that pulls in a large file both count as "a message," but the tokens involved, and so the credits spent, can differ by an order of magnitude. Cached context lowers the input-token cost; a longer reply raises the output-token cost.