LLM APIs bill per million tokens on separate meters: new input, cached input and output. Reasoning tokens bill as output, images bill as input tokens, and speech to text bills per audio minute. This guide explains each meter, then prices real requests with BroadNet Inference’s published rates, synced 2026-11-14.
What are you actually paying for?
A token is a chunk of text, usually a few characters. In BroadRouter’s own measurement across five tokenizers, 1,000 English words came to about 1,290 tokens; OpenAI’s rule of thumb is about 75 words per 100 tokens. Other languages differ; paste your own text into the free token counter to see the real number. Each request reports its tokens in the usage block, and the bill is the sum of these meters:
| Meter | What it counts | Billed at |
|---|---|---|
| Input (prompt) | System prompt, history, documents, tool definitions, images | Input price |
| Cached input | Input the provider already processed and reused | Cached price, much lower |
| Output (completion) | The visible answer plus any reasoning | Output price |
| Audio minutes | Length of audio sent for transcription | Price per minute |
The input and output token split matters most, because the two prices are far apart.
What does each model cost per token?
BroadNet Inference prices, per 1M tokens:
| Model | Input | Cached input | Output | Context |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | $0.175 | $0.0055 | $0.52 | 1,048,576 |
| GLM 5.3 | $0.78 | $0.173 | $3.46 | 1,048,576 |
| Qwen3.8 27B | $0.175 | $0.0435 | $2.17 | 262,144 |
| Gemma 4 VL 26B | $0.10 | No cache discount | $0.22 | 32,768 |
Two patterns hold on this rate card and on most others. Output costs more than input, from about twice as much on Gemma 4 VL 26B to over ten times on Qwen3.8 27B. Cached input costs a fraction of new input. To compare these rates with OpenAI, Anthropic, Google and other providers, see the LLM API pricing table. For market-wide medians, provider spread and output-to-input ratios across every model we track, see the LLM API price index (open CSV and JSON).
Why is output priced higher than input?
The model reads your whole prompt in one parallel pass. It writes the answer one token at a time, and each new token needs another pass. Output therefore uses more GPU time per token, and the price reflects it. A short prompt with a long answer can cost more than a long prompt with a short answer.
How are cached input tokens billed?
When the start of your prompt matches a recent request, the provider can reuse its work. Those cached input tokens appear as prompt_tokens_details.cached_tokens and bill at the cached price. On DeepSeek V4.1 Flash that is $0.0055 instead of $0.175 per 1M.
Caching rewards prompts with a stable prefix: system prompt, tool list and documents first, the new message last. Agents that resend the same repository on every turn gain the most. For how each provider bills cache reads and writes, see cached input tokens explained; to model your own cache rate, use the prompt caching calculator.
Why does a reasoning model bill more output than you see?
DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B think before they answer. That thinking is made of reasoning tokens. You don’t see them in the reply, but they are counted in completion_tokens and billed at the output price. OpenAI’s reasoning guide describes the same rule for its models: reasoning tokens “are billed as output tokens”.
Reasoning also counts toward max_tokens. Set it too low and you pay for thinking with no answer. Why replies come back empty, and how to keep reasoning costs down.
How are images and audio priced?
Images. A vision model turns an image into input tokens. You pay the input price for them, and a bigger image uses more tokens. There is no separate per-image fee on BroadNet. Cost per image compared.
Audio. Speech to text is billed by audio length, not tokens:
| Model | Per minute | Per hour of audio |
|---|---|---|
| BroadNet Transcribe Arabic | $0.0042 | $0.25 |
| BroadNet Transcribe (14 languages) | $0.002 | $0.12 |
What does a real request cost? Worked examples
These use DeepSeek V4.1 Flash. Each example is sized so the totals land on whole millions of tokens, which makes the math easy to check.
- A chat feature. 500 requests, each with 2,000 input and 2,000 output tokens, use 1M input and 1M output tokens. Cost: $0.175 plus $0.52.
- An agent rereading its context. Ten turns that each send the same 100,000-token context add up to 1M input tokens. If none of it is cached, that costs $0.175. If it were all cached, it would cost $0.0055, but the first turn has nothing to reuse, so treat that as a lower bound: at best, 900,000 tokens bill at the cached price and the first 100,000 at the input price.
- Hidden reasoning. 1,000 answers of 200 visible tokens, each with 800 reasoning tokens, bill 1M output tokens: $0.52. Only a fifth of that was text you saw.
- Transcription. 1,000 hours of Arabic audio cost $252.00. 10,000 voice notes of 30 seconds cost $21.00.
The general formula:
cost = new_input/1M × input_price
+ cached_input/1M × cached_price
+ output/1M × output_price (output includes reasoning)
To run these numbers for your own traffic, use the free LLM API cost calculator. For a full forecast of a new feature, see how to estimate LLM API cost before launch.
How do you read the real numbers from a response?
Don’t estimate from word counts once you have traffic. Read them:
import os
from openai import OpenAI
client = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])
r = client.chat.completions.create(
model="broadnet/deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Summarise this ticket in one line: ..."}],
)
u = r.usage
details = u.prompt_tokens_details # can be None when the model reports no cache data
cached = (details.cached_tokens or 0) if details else 0
print(u.prompt_tokens, cached, u.completion_tokens)
When streaming, add stream_options={"include_usage": True} so the last chunk carries usage. Streaming guide.
What other pricing models exist?
Per-token billing is not the only option:
- Flat plans. A flat-rate plan charges for time, not tokens. BroadNet’s Full Moon is $60 a month per session for unlimited DeepSeek V4.1 Flash, one request at a time. Orbit is $15 a week, Meteor $1 an hour.
- Hourly keys. $0.50 buys one hour of Gemma 4 VL 26B on one stream. Hourly keys explained.
- Batch and dedicated GPUs. Some providers discount overnight batch jobs or rent whole GPUs by the hour. How other providers price open models.
When a flat plan beats per-token depends on volume and cache rate. For every lever in order, from caching to batch discounts, see how to reduce LLM API costs.
Per-token prices also change over time, in both directions. In archived pricing pages from 11 providers, 80% of models never changed price. Across the models that did change, we recorded 100 price cuts and 52 rises, each a model’s price compared with its own earlier price. See how often LLM API prices change and the biggest LLM API price cuts, with sources.
Where BroadNet Inference fits
BroadNet publishes one rate card for every model it serves, in a machine-readable file that this site’s pricing page reads. New accounts start with $5 of credit; Gemma 4 VL 26B and both transcription models work on it, and the other chat models need a top-up. It is a good fit if you want per-token and flat pricing on one account; plans get their own keys. It is not the place for frontier closed models: it serves open-weight models and its own builds only.

