Running Codex with GPT-5.6 Sol on the API, each step (60K context tokens, 90% cached, 800 output) costs $0.0810; a 40-step session, $3.24. At 4 sessions a day over 22 days: $285.12 a month.
On a plan, the agent is covered by the fixed price within usage limits; on the API you pay every token. People who code with an agent all day usually save with a Max plan; CI automation fits the API better.
Every step resends instructions, history and the files read so far. Without caching the bill would be several times higher, which is why short, focused sessions cost less.
Start fresh sessions for new tasks, keep the instructions file short and use a smaller model for mechanical work.
No. Claude and ChatGPT cap usage per time window (on Claude, 5-hour windows plus weekly limits). Bigger plans give more usage.
Estimate based on public prices as of October 2026. Token counts vary with the text and the model; check the official pricing page before you decide.