Every LLM request has two sides: what you send and what comes back. APIs count both in tokens, and most bill them at different rates.
How it works
Input is everything in the request: the system prompt, every earlier message, tool definitions, tool results, documents and images. A multi-turn chat resends the whole history each turn, so input grows with the conversation.
Output is everything the model generates: the visible answer, tool calls and, on reasoning models, the hidden reasoning tokens.
The usage block reports both:
"usage": {
"prompt_tokens": 18000,
"completion_tokens": 900,
"prompt_tokens_details": { "cached_tokens": 15000 },
"completion_tokens_details": { "reasoning_tokens": 600 }
}
The cost of one request is:
(new input × input price) + (cached input × cached price) + (output × output price), with each price per 1M tokens.
Why it matters
The ratio between the two decides which price matters. A chat assistant may produce a few hundred tokens for a short question. A coding agent resends a large repository context on each turn and writes a short patch, so input can outnumber output by 20 to 1 or more. For agents, the input price and the cached input price drive most of the bill.
Output is still worth watching. Reasoning models can spend more tokens thinking than answering, and those are billed as output. Across the models we track, the typical output-to-input price ratio is in the LLM API price index.
On BroadNet Inference
Per 1M tokens, from the published price file:
| Model | Input | Cached input | Output |
|---|---|---|---|
| DeepSeek V4.1 Flash | $0.175 | $0.0055 | $0.52 |
| GLM 5.3 | $0.78 | $0.173 | $3.46 |
| Qwen3.8 27B | $0.175 | $0.0435 | $2.17 |
| BroadNet Gemma 4 VL 26B | $0.10 | — | $0.22 |
In BroadNet’s worked example, a full-time coding agent sends 1,000 million input tokens a month with 80% cache hits and gets 50 million output tokens back. On DeepSeek V4.1 Flash per token, that costs $65.40. A flat-rate plan removes the meter: Full Moon is $60 a month per session. Compare your own numbers in the pricing calculator or read flat rate vs pay per token.
Common mistakes
- Pricing only the visible answer and forgetting reasoning tokens.
- Ignoring how fast history grows. Turn 30 of a chat can carry 30 turns of input.