Reasoning models think before they answer. That thinking is text too, so it costs tokens, time and space in the context window.
How it works
A reasoning model generates an internal chain of thought, then the visible answer. Both come from the same output budget. OpenAI’s guide says reasoning tokens “are billed as output tokens” and occupy space in the context window; Anthropic says the same of thinking tokens and calls them “a subset of your max_tokens.”
In the Chat Completions format, the count appears in usage:
"usage": {
"completion_tokens": 1186,
"completion_tokens_details": { "reasoning_tokens": 1024 }
}
Here, 1,024 of 1,186 output tokens went to thinking, and only 162 to the answer. OpenAI warns that a response can hit the limit “before any visible output tokens are produced,” so you pay for input and reasoning and get no reply. It suggests reserving at least 25,000 tokens for reasoning and output when you start experimenting with its models.
Many APIs expose a control such as reasoning_effort (low, medium, high) to trade quality for fewer thinking tokens.
Why it matters
- Cost. On a per-token key, reasoning is billed at the output price, usually the highest rate. See input and output tokens.
- Latency. Thinking takes time before the first visible token. At a fixed tokens per second, 2,000 reasoning tokens add the time of 2,000 output tokens.
- Reliability. A small
max_tokensthat worked on a non-reasoning model can return empty answers on a reasoning one.
On BroadNet Inference
DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B reason before they answer. BroadNet’s docs say the reasoning counts toward max_tokens. BroadNet’s docs do not list reasoning for BroadNet Gemma 4 VL 26B.
When the model reports it, usage.completion_tokens_details.reasoning_tokens shows the thinking share. Set max_tokens to 1,000 or more, or leave it unset, and use reasoning_effort="low" for simple steps:
client.chat.completions.create(
model="broadnet/deepseek-v4.1-flash",
reasoning_effort="low",
max_tokens=1000,
messages=[{"role": "user", "content": "Classify: 'refund my order'"}],
)
On pay-as-you-go, reasoning is billed at the output price, for example $0.52 per 1M tokens on DeepSeek V4.1 Flash. On an unlimited plan for that model there is no token meter, so reasoning costs nothing extra; it only takes time. The full fix list is in why a reasoning model returns an empty answer.