You send a short question, get a two-line answer, and the bill shows 1,400 output tokens. Nothing is wrong. The model thought before it answered, and the thinking was billed. This post explains how reasoning tokens are counted, what they cost on each model, and the four settings that control them.
What are reasoning tokens?
A reasoning model writes an internal chain of thought, then the visible answer. Both come out of the same generation. The thinking is made of reasoning tokens. In the standard chat-completions response the answer is in message.content, and the thinking is counted, not shown.
OpenAI’s reasoning guide states the rule plainly: reasoning tokens “still occupy space in the model’s context window and are billed as” output tokens. Other providers follow the same rule.
How are reasoning tokens billed?
They are part of completion_tokens, so they pay the output price. Output is the expensive meter:
| Model | Reasons first | Output per 1M | Input per 1M |
|---|---|---|---|
| DeepSeek V4.1 Flash | Yes | $0.52 | $0.175 |
| GLM 5.3 | Yes | $3.46 | $0.78 |
| Qwen3.8 27B | Yes | $2.17 | $0.175 |
| Gemma 4 VL 26B | Not listed | $0.22 | $0.10 |
Prices are BroadNet Inference’s, synced 2026-11-14. What other hosts charge for the same models is in the LLM API pricing table, for example GLM 5.3 across providers.
How much can reasoning add to a bill?
Take a classifier that returns a one-word label, about 10 visible tokens. If the model thinks for 1,000 tokens first, 99% of the output bill is thinking.
At scale: 1,000 replies with 1,000 reasoning tokens each add 1M output tokens. That costs $0.52 on DeepSeek V4.1 Flash, $2.17 on Qwen3.8 27B and $3.46 on GLM 5.3, on top of the visible answers.
The amount varies by task. Hard maths or multi-file code reasoning can run to thousands of tokens. A greeting may need almost none. That is why you measure.
How do you measure reasoning tokens?
The usage block reports them when the model provides the count:
r = client.chat.completions.create(
model="broadnet/deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Is 'refund my order' a billing or shipping intent?"}],
max_tokens=1000,
)
u = r.usage
reasoning = (u.completion_tokens_details.reasoning_tokens or 0) if u.completion_tokens_details else 0
print("output:", u.completion_tokens, "reasoning:", reasoning,
"share:", round(reasoning / max(u.completion_tokens, 1), 2))
Log this per task type for a week. You will usually find a few tasks responsible for most of the thinking. When streaming, add stream_options: {"include_usage": true}; the counts arrive in the last chunk (streaming guide).
Four ways to control reasoning cost
1. Lower reasoning_effort for simple tasks
reasoning_effort asks the model to think less. BroadNet’s guides use low for classification, extraction and formatting:
client.chat.completions.create(model="broadnet/deepseek-v4.1-flash", reasoning_effort="low",
max_tokens=1000, messages=msgs)
OpenAI documents a wider range of values on its own models, from none and minimal through low, medium and high to xhigh and max, and notes that “some models support only a subset”. Check which values your model accepts, and compare output tokens before and after on the same samples.
2. Set max_tokens as a ceiling, not a squeeze
max_tokens caps reasoning and answer together. It stops a runaway reply, but if you set it too low, the model can spend the whole budget thinking and return an empty answer with finish_reason: "length". You pay for that thinking and still have to retry. BroadNet suggests 1,000 tokens or more for reasoning models. Why a reasoning model returns an empty answer.
3. Route easy steps to a model that does not reason
Not every step needs thought. Formatting JSON, renaming fields or answering FAQ lookups can go to Gemma 4 VL 26B at $0.22 per 1M output, for which BroadNet does not list reasoning. Keep the reasoning models for planning, debugging and multi-step answers. Test quality on your own samples before routing.
4. Remove “think step by step” from prompts
Reasoning models already think. Asking for visible step-by-step reasoning on top makes them write it twice: once hidden, once in the answer. Ask for the final answer in the format you need.
Which setting should you use for which task?
| Task | Reasoning effort | max_tokens | Model idea |
|---|---|---|---|
| Classification, routing | low | 1,000 | Gemma 4 VL 26B or DeepSeek V4.1 Flash |
| Extraction to JSON | low | 1,000 to 2,000 | DeepSeek V4.1 Flash, GLM 5.3 for structured output |
| Customer reply drafting | low or default | 2,000 | DeepSeek V4.1 Flash |
| Debugging, planning, code review | default or higher | 8,000 or unset | DeepSeek V4.1 Flash or GLM 5.3 |
These are starting points. Measure output tokens and quality on your own data, then adjust.
When does reasoning stop costing extra?
On a flat plan. BroadNet’s Full Moon plan is $60 a month per session for unlimited DeepSeek V4.1 Flash with no token meter, so reasoning tokens have no per-token cost. They still take time: at a typical speed of about 100 tokens per second, every reasoning token is generated before the answer starts, so long thinking delays the reply. When a flat plan beats per-token, and ten more ways to cut a coding agent’s bill.
Where BroadNet Inference fits
BroadNet runs three reasoning models and one model with no listed reasoning on one OpenAI-compatible key at https://inference.broadnet.ai/v1, with usage reported in each response. That makes it straightforward to route by task and measure the difference. If you need a closed model with a visible reasoning summary, use that vendor’s API; BroadNet’s chat catalogue has no closed models. See the plans.
