See what your model costs.
Compare model rates. Explore how more reasoning changes usage.
Compare models
Compare reasoning
Effort levels have no fixed cost multiplier. These bars use the editable token assumptions below.
Your sample workload Adjust the assumptions
Starting values are illustrative, not measured averages or OpenAI token budgets. More reasoning uses output tokens; the actual amount varies by task.
Assumed reasoning tokens per effort
Ultra: total cost can include multiple agents. Enter total uncached input, cached input, answer and reasoning tokens across all agents, assuming the selected model for every agent. Leave Ultra reasoning blank to keep its cost unknown. Mixed-model teams need separate estimates.
Rates, sources & what this includes
| Model | Input | Cached input | Output |
|---|
Credits compare model consumption; they do not convert directly to your subscription’s remaining weekly percentage or a dollar charge. Tools, images, voice, retries and additional turns can add usage. Rates are a dated snapshot and do not refresh automatically.
Calculation: (uncached input × input rate + cached input × cached rate + (answer + reasoning) × output rate) ÷ 1,000,000. Cached input is separate from uncached input; output includes reasoning.
Effort availability reflects your local Codex model catalog on September 21, 2026. Light is called Low in the CLI. Luna has no Ultra; GPT-5.5 has no Max or Ultra. GPT-5.5 retires from Codex with ChatGPT sign-in on October 14, 2026.
Official Codex pricing ↗ · Models & effort guidance ↗ · Reasoning token accounting ↗