On the public API price lists checked on 19 October 2026, output tokens cost five or six times as much as input tokens. Anthropic and OpenAI list a 5 to 1 ratio on most current models, and Google lists 5 to 1 or 6 to 1. Long answers cost more than long questions.
The ratios, with dates
All prices below are US dollars per million tokens for standard processing, read from each provider's public pricing page on 19 October 2026. Prices change, so treat this as a snapshot and check the source before you budget.
| Provider and model | Input | Output | Ratio |
|---|---|---|---|
| Anthropic, Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.50 | 5 to 1 |
| Anthropic, Sonnet 5.5 | $2 | $10 | 5 to 1 |
| Anthropic, Opus 5.5 | $4 | $20 | 5 to 1 |
| OpenAI, gpt-6-luna | $0.10 | $0.50 | 5 to 1 |
| OpenAI, gpt-6.1-sol | $2.00 | $10.00 | 5 to 1 |
| Google, Gemini 3.1 Flash-Lite | $0.25 | $1.50 | 6 to 1 |
| Google, Gemini 3.1 Pro Preview (up to 200k) | $2.00 | $12.00 | 6 to 1 |
The pages do not say why the ratio is what it is, and this post will not guess at a provider's internal costs. What matters for a buyer is the shape of the list.
What this means for your workload
Whether input or output dominates your bill depends on the job.
- Summarizing a long document: input heavy. A 20,000 token report in and a 500 token summary out means input is most of the cost even at a 5 to 1 ratio.
- Writing code or long drafts: output heavy. Each response may be thousands of tokens.
- Chat with history: input grows every turn because the history is resent, so input usually wins in long sessions.
- Extended thinking: billed as output on Anthropic models, so a hard question can produce a large output bill before any visible answer appears.
Cheap levers on each side
To reduce input cost, trim what you send, cache stable prefixes, and avoid pasting whole files when a section will do. See prompt caching explained.
To reduce output cost, ask for the format you need, such as a short answer or a diff, set a sensible maximum, and lower the reasoning effort on simple tasks where the model allows it.
Do not compare list prices alone
A cheaper model that needs three attempts can cost more than a pricier one that gets it right once. Compare cost per finished task on your own examples. Choosing a model tier by cost and task covers a simple method.
Quick check
Is output always five times the price of input?
Not always. On the pages read for this post the ratio was 5 to 1 or 6 to 1 for the models listed, but other models, modes and long context tiers can differ.
Where do I find my real input and output split?
In the usage object each API response returns, and in your provider's usage dashboard. Log it per feature.
Sources
Last reviewed: 19 October 2026.
