🧠 AI Tip: Hidden thinking tokens are billed, even when you never see them
OpenAI's reasoning models plan before they answer. Those tokens never appear in the API response, but they are still billed as output.
The dial is reasoning.effort: none, minimal, low, medium, high, xhigh, or max, depending on the model. Lower is faster and cheaper. gpt-5.5 defaults to medium. So do GPT-5.6, GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna if you leave it unset.
Use low for support replies, drafts, and simple code. Keep high for hard debugging, and xhigh only if your own evals say so. GPT-6 Astra rejects "none" with HTTP 400. GPT-6.1 Sol rejects none and minimal.
A max_output_tokens cap that is too small can be spent entirely on thinking. You get "incomplete", no answer, and still a bill. While testing, reserve at least 25,000 tokens for reasoning plus the answer.
Example: {"effort": "low"} for a support reply, "high" for a hard debug.
⭐ Match the dial to the job.
#AI
OpenAI's reasoning models plan before they answer. Those tokens never appear in the API response, but they are still billed as output.
The dial is reasoning.effort: none, minimal, low, medium, high, xhigh, or max, depending on the model. Lower is faster and cheaper. gpt-5.5 defaults to medium. So do GPT-5.6, GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna if you leave it unset.
Use low for support replies, drafts, and simple code. Keep high for hard debugging, and xhigh only if your own evals say so. GPT-6 Astra rejects "none" with HTTP 400. GPT-6.1 Sol rejects none and minimal.
A max_output_tokens cap that is too small can be spent entirely on thinking. You get "incomplete", no answer, and still a bill. While testing, reserve at least 25,000 tokens for reasoning plus the answer.
Example: {"effort": "low"} for a support reply, "high" for a hard debug.
⭐ Match the dial to the job.
#AI
