Why output tokens cost more than input tokens
On nearly every model in this index, generating a token costs more than reading one. That ratio quietly decides which features are affordable.
Read a pricing page and the first surprise is that input and output are billed at different rates, with output almost always the more expensive of the two. This is not an accounting trick. Producing a token is a different operation from consuming one, and the price difference is stable enough that it should be treated as a product constraint.
What the ratio looks like in this dataset
Across 1509 priced chat models that publish both rates, the median output rate is $1.5 per 1M tokens and the median input rate $0.6 — a median ratio of about 3.4×. But the spread around that median is the interesting part:
| Model | Provider | Input /1M | Output /1M | Ratio |
|---|---|---|---|---|
| Qwen3 Vl 32b Thinking | Alibaba DashScope | $0.16 | $2.87 | 17.9× |
| Nemotron 3.5 Lightning 30b A3b | Perplexity | $0.011 | $0.17 | 14.8× |
| Llama 3.2 11b Vision Instruct | Cloudflare Workers AI | $0.049 | $0.676 | 13.9× |
| … | … | … | … | … |
| Llama Guard 3 8b | Cloudflare Workers AI | $0.484 | $0.03 | 0.1× |
| Qwen2 1.5B Instruct | Together AI | $0.02 | $0.02 | 1.0× |
| Llama 3.2 3B Instruct | Together AI | $0.06 | $0.06 | 1.0× |
Why the ratio matters more than the headline rate
- Verbose features pay the ratio. A report generator that emits 4,000 tokens per request is buying the expensive kind of token in bulk. Two models with identical input prices can differ by a factor of two on that workload.
- Chat pays it less than you think. A support answer of 150 tokens against 3,000 input tokens is input-dominated; for that shape, input price and cache rates decide the bill and the output column is close to noise.
- Agent loops pay it worst. Tool-calling loops resend the transcript as input and generate several intermediate completions, so they buy both sides of the ratio at volume.
The cache caveat
A cached input token is the cheapest token on any pricing page. That is why a high output/input ratio is not automatically a bad sign: if your workload is 80% stable prefix, the input side of the ratio barely applies to you. Compare the cached column first, then the ratio.
Pick the workload shape closest to your product and see the ranking change.
Browse the workload rankingsFrequently asked
› What is a typical output-to-input price ratio?
Across the priced chat models in this index the median output rate is $1.5 per 1M tokens against a median input rate of $0.6 — a ratio of roughly 2.5×. Individual models vary widely in both directions, which is why the ratio is worth checking per model rather than assuming.
› Why are cached input tokens cheaper than fresh input?
A cached prefix is not re-processed the same way fresh input is, so providers publish a separate, lower rate for it. It is the only discount in this market that applies to work you were already doing.
› Should I choose models by the ratio?
Only as a tiebreaker for a specific workload shape. A verbose product (long reports, code generation, agent loops) should weight output price heavily; a retrieval-heavy product should weight input price and cache rates.
Where the numbers come from
The data pages this article draws on — check any figure against them, they are recomputed on every refresh.