Why output tokens cost more than input tokens

Look at any price list

Pick almost any model and the output price is a multiple of the input price. Here are five from the calculator:

Standard-tier prices per million tokens, as listed in the calculator.

The gap isn't a pricing quirk. It comes from how models produce text.

Reading is parallel, writing is one token at a time

When a model reads your prompt, it processes all of the input tokens together in a single pass. This step is called prefill. GPUs are built for exactly this kind of large, parallel calculation, so each input token is cheap to process.

Output works differently. The model generates one token, adds it to the text, then runs again to produce the next. This step is called decode. Each new token needs its own pass through the model and can't start until the previous one is finished. For most of that pass, the GPU is waiting on memory, moving model weights and the cached context around, rather than calculating.

So an output token ties up expensive hardware for far longer than an input token does. Providers price that difference in.

What it means for your bill

Take the short email example on the calculator's homepage: about 112 input tokens and 123 output tokens. On Claude Sonnet 5.5:

• Input: 112 tokens × $2 per million = $0.000224
• Output: 123 tokens × $10 per million = $0.00123

The token counts are nearly equal, but output is 85% of the cost. Send 10,000 of those emails and you pay $14.54, of which $12.30 is output.

Four ways to spend less on output

1. Ask for shorter answers. State the length you want, such as "in 3 bullet points" or "under 100 words", and set a maximum output token limit so one long answer can't run up the bill.
2. Ask for structured output. JSON or a fixed template usually takes fewer tokens than flowing prose, and it's easier for your code to read.
3. Use batch processing when you can wait. Anthropic, OpenAI and Amazon charge 50% less for batch jobs that return results later instead of instantly.
4. Match the model to the task. Small models such as Claude Haiku 5.5 or GPT-4.1-nano charge $0.40 to $0.50 per million output tokens, compared with $8 to $10 for mid-size models. For simple tasks the quality difference is often small.