Why output tokens cost more than input tokens
Every major AI API charges more for the words a model writes than for the words it reads. This page measures how much more across the models in our price table, shows how that changes what your bill is made of, and explains the likely technical reasons, clearly labelled as our explanation. Price your own workload in the cost calculator; the wider topic is AI costs and pricing.
Prices last verified against official provider pages on . How we verify prices. Written by the Chelure Enterprises Editorial Team.
The ratio across providers
Ratio = output price divided by input price, per million tokens, using the current list prices in our price comparison.
| Provider | Models | Lowest ratio | Median | Highest ratio |
|---|---|---|---|---|
| Anthropic | 11 | 5.0x | 5.0x | 5.0x |
| DeepSeek | 2 | 3.0x | 3.5x | 4.0x |
| 8 | 4.0x | 5.0x | 8.3x | |
| Mistral | 6 | 1.0x | 2.0x | 4.0x |
| Moonshot | 6 | 4.2x | 4.6x | 5.0x |
| OpenAI | 17 | 4.0x | 6.0x | 8.0x |
| Perplexity | 1 | 10.0x | 10.0x | 10.0x |
| Z.ai | 8 | 3.1x | 3.3x | 5.5x |
| xAI | 6 | 2.0x | 2.5x | 3.0x |
Across all 65 models the ratio runs from 1.0x to 10.0x with a median of 5.0x.
What this does to your bill
Because output is priced several times higher, it can dominate the cost even when it is a small share of the tokens. On Claude Sonnet 5.5 ($2 input, $10 output per million tokens):
| Workload | Input / output tokens | Output share of tokens | Output share of cost | Cost per request |
|---|---|---|---|---|
| Chat turn | 1,500 / 400 | 21% | 57% | $0.0070 |
| Summarising a document | 8,000 / 400 | 5% | 20% | $0.02 |
| Long answer from a short prompt | 300 / 1,500 | 83% | 96% | $0.016 |
| Classification (one-word reply) | 600 / 5 | 1% | 4% | $0.0012 |
A chat turn is about 21 percent output by tokens but over half the cost; a short-prompt, long-answer task is almost entirely output cost. That is why trimming replies usually saves more than trimming prompts, and why prompt caching, which only discounts input, helps most when your prompts are long and your answers are short.
Why providers price it this way (our explanation)
Provider pricing pages state the prices and not the cost reasoning, so this section is our interpretation. In the Transformer architecture introduced by Vaswani et al. (2017), a model can process all the tokens of a prompt in one pass, but it produces output one token at a time, each depending on the ones before. Kwon et al. (2023) describe LLM serving throughput as limited by memory: high throughput needs large batches, but each request's key-value cache is large and grows as it generates, which limits how many requests fit. A long answer therefore occupies serving capacity for longer than reading a long prompt does. Pricing output higher passes that cost on.
Practical ways to cut the output share
- Ask for concise answers and set
max_tokensto a sensible ceiling. - Return structured fields rather than prose.
- Route simple tasks to a cheaper model; see Haiku pricing for an example.
- Use batch processing for work that can wait: it halves both input and output.
Frequently asked questions
Why do output tokens cost more than input tokens?
The providers' pages we read list the prices but do not explain the reason. The usual technical explanation, which is ours and not a provider statement, is that input tokens are processed together in parallel while output tokens are generated one at a time, and each request holds memory for its context while it generates. Research on LLM serving describes memory for the key-value cache as what limits how many requests can be batched.
How much more do output tokens cost?
Across the 65 models in our price table the output price is between 1.0 and 10.0 times the input price, with a median of 5.0 times.
Do all providers use the same ratio?
No. The ratio ranges from about 2 times for some models to 5 times or more for others; the table on this page shows the range for each provider.
How do I reduce output costs?
Ask for shorter answers, use structured output with only the fields you need, set a sensible max_tokens, and use smaller models for simple tasks. Reasoning or thinking models can produce extra tokens billed as output on some providers, so check their pricing pages.
Does the ratio change for cached input?
Cached input is cheaper than normal input, which widens the gap further: for a mostly cached prompt, the output side dominates the bill.
Sources
- Anthropic pricing
- OpenAI pricing
- Google Gemini pricing
- Vaswani et al.: Attention Is All You Need (arXiv:1706.03762)
- Kwon et al.: Efficient Memory Management for LLM Serving with PagedAttention (arXiv:2309.06180)
Primary sources opened on November 13, 2026. Prices come from our verified price table (see the methodology).