Output tokens cost more than input tokens on every model in my price list, so the shape of a task decides its bill. A large document with a short answer (summarizing a PDF, RAG) is dominated by input. A short prompt with a long answer (a blog post) is dominated by output, and by reasoning tokens, which are billed at the output rate.
Why output costs more
Every model on TaskCents lists separate prices for input, output and cached input. Here are some pairs from the price list (USD per 1M tokens, as of 2026-10-11):
| Model | Input | Output | Cached input |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 |
| Kimi K3 | $0.80 | $15.00 | $0.55 |
| GLM 5.3 | $0.039 | $4.80 | $0.038 |
For the Anthropic and OpenAI models above, output is five times the input price. Kimi K3 and GLM 5.3 have far larger gaps. Even the narrowest gaps in the list, such as MiMo-V2.6-Pro ($0.43 input, $0.87 output), still price output higher.
Two things follow:
- A long document is cheap to read but expensive to write about.
- The cheapest way to lower an output-heavy bill is usually to shorten the answer, not the prompt.
Input-heavy work: summarizing a PDF and RAG
My summarize-PDF task uses a 20-page PDF of 10,023 words (J. S. Mill's On Liberty). The model reads the whole text and returns a short summary.
Claude Sonnet 5.5 used 18,004 input tokens, 857 output tokens and no reasoning tokens. At $2.00 input and $10.00 output per million, input is roughly 80% of the 4.5¢ total.
GPT-6 Luna used 12,987 input tokens, 449 output tokens and 84 reasoning tokens. At $0.10 and $0.50, the total is 0.16¢, and about 83% of that is input.
The RAG task (an answer from 5 retrieved documents) splits differently. Claude Sonnet 5.5 used 5,184 input tokens, 1,079 output tokens and 615 reasoning tokens, for 2.7¢. Input is about 38% of that cost. A short context plus a reasoning step moves the bill toward the output side.
Tokenizers matter here too. Anthropic's tokenizer produces 1.39 tokens for the same English text where OpenAI's produces 1.00. That is why Claude Sonnet 5.5 and Claude Haiku 5.5 read 18,004 tokens of the PDF while GPT-6 Luna read 12,987.
Output-heavy work: a 1,500-word blog post
The blog post task has a very short prompt (52 to 104 input tokens for most models) and a long answer. Claude Sonnet 5.5 used 90 input tokens, 3,401 output tokens and 239 reasoning tokens. That is about four times the output of its PDF summary.
Input is a tiny fraction of the 3.7¢ total. Here the bill is almost entirely output. The answer's length and the model's reasoning habits matter far more than the prompt.
The same models can rank very differently across the two tasks:
| Model | Input / Output ($ per 1M) | Summarize PDF | Blog post | Basis |
|---|---|---|---|---|
| GPT-6 Luna | $0.10 / $0.50 | 0.16¢ | 0.12¢ | measured |
| Claude Haiku 5.5 | $0.10 / $0.50 | 0.22¢ | 0.19¢ | measured |
| Claude Sonnet 5.5 | $2.00 / $10.00 | 4.5¢ | 3.7¢ | measured |
| GLM 5.3 | $0.039 / $4.80 | 0.92¢ | 3.3¢ | estimated |
| Kimi K3 | $0.80 / $15.00 | 4.3¢ | 12¢ | estimated |
Kimi K3 shows the pattern best. The PDF task used 13,113 input tokens and cost 4.3¢. The blog post used only 53 input tokens but cost 12¢, because the model produced 2,023 output tokens and 5,705 reasoning tokens. GLM 5.3 moves the same way, from 0.92¢ to 3.3¢. A very low input price does not help when the answer is long.
Some runs on the blog task never produced visible text. Qwen3.8 Flash (measured), Llama 4 Maverick and Qwen3.8 Max Prime (both estimated) hit the 16,000-token output cap with 0 visible output tokens, all spent on reasoning. The Qwen3.8 Max Prime estimate for that task is 19¢.
Reasoning tokens bill as output
Reasoning tokens are billed at the output rate. This matters most for models whose reasoning mode is "default" or "always".
In the summarize task, Qwen3.8 Flash used 2,141 reasoning tokens against 501 visible output tokens. Most of its billed output was invisible. On the blog task, GLM 5.3 Flash used 4,749 reasoning tokens against 2,174 output tokens.
When you compare models, read the reasoning column as well as the answer length. A model with a low output price can still cost more than expected if it thinks at length. My calculator lets you enter your own token counts, and the methodology page explains how I measure tokens and which runs are estimated.
What to do with this
- For document-heavy tasks, compare models on input price first, then check output and reasoning.
- For writing tasks, compare output prices and typical answer length. A low input price will not help much.
- Cap output where the task allows it. Shorter answers usually save more than shorter prompts.
- Check reasoning settings before you commit. Some models in the list have reasoning mode "none", which removes that output cost.
Model-level detail is on the Claude Sonnet 5.5 and GPT-6 Luna pages. For plan options, see plans.
FAQ
Why are output tokens priced higher than input tokens?
Providers set the prices, and in my list output is priced above input for every model shown. For the Anthropic and OpenAI models in this article, output costs five times as much as input.
Do reasoning tokens count toward the output price?
Yes. Reasoning tokens are billed as output tokens. A model that reasons heavily can cost more than its visible answer suggests, as the Qwen3.8 Flash summary run shows.
Is a long prompt always cheap?
Not always. A long prompt is cheap on models with low input prices when the answer is short. It gets expensive on models with high input prices. Check both prices and the typical output length before you decide.