AI models now charge five times more to write than to read
Every model has two prices: one for the text it reads and one for the text it writes. Among models from late 2024 the writing price was typically three times the reading price. Among this year's models it is five.
Price lists for AI models quote two numbers per million tokens. The input price covers everything you send: your question, your documents, the conversation so far. The output price covers what the model writes back. For most models output costs more. The gap between the two has widened.
We divided each model's output price by its input price and took the median for models released in each half-year. For models from the second half of 2024 the median is 3. For models from the second half of 2025 onwards it is 5.
Median output price as a multiple of input price, by release period
| Released | Models | Median ratio | Charging 5x or more |
|---|---|---|---|
| 2024 H2 | 31 | 3x | 5 (16%) |
| 2025 H1 | 47 | 4x | 10 (21%) |
| 2025 H2 | 49 | 5x | 28 (57%) |
| 2026 H1 | 60 | 5x | 35 (58%) |
| 2026 H2 | 51 | 5x | 36 (71%) |
The median has sat at five since the second half of 2025, but the last column kept moving. In late 2024 about one model in six charged five times or more for output. Since July this year it is seven in ten.
Five has become the house number
Among the 65 models we list as active and priced, 34 charge five times or more for output, and 14 charge exactly five.
Active language models by output-to-input price ratio
Anthropic is the plainest case. All 10 of its models in this set are priced at exactly five to one, from Claude Haiku 4.5 at $1 in and $5 out to Claude Fable 5 at $10 and $50. OpenAI's current models sit a little higher, at about six. The lowest ratio in the set is Llama 3.3 70B Instruct, which charges the same for both, and the highest is Gemini 3.5 Flash-Lite at 8.3 to one.
| Model | Input | Output | Ratio |
|---|---|---|---|
| Mistral Nemo (Mistral) | $0.029 | $0.03 | 1.0x |
| Llama 3.3 70B Instruct (Meta) | $0.71 | $0.71 | 1.0x |
| Claude Haiku 4.5 (Anthropic) | $1 | $5 | 5.0x |
| Claude Fable 5 (Anthropic) | $10 | $50 | 5.0x |
| GPT-5.5 (OpenAI) | $5 | $30 | 6.0x |
| GPT-5.5 Pro (OpenAI) | $30 | $180 | 6.0x |
| GPT-5.4 Nano (OpenAI) | $0.2 | $1.25 | 6.3x |
| Gemini 3.5 Flash-Lite (Google) | $0.3 | $2.5 | 8.3x |
The models near one to one are small open-weight models sold by third-party hosts. The models at five and six are the labs' own flagships. We can see that pattern in the prices. We cannot see the labs' costs, so we are not going to guess at why they price this way.
Which work this makes dearer
A higher output price only bites when a job writes a lot. Here are three jobs priced at GPT-5.5's list price of $5 per million input tokens and $30 per million output tokens. The token counts are made-up round numbers to show the arithmetic, not measurements.
| Job | Tokens in | Tokens out | Input cost | Output cost | Output share |
|---|---|---|---|---|---|
| Summarise a long report Reads a lot, writes a little | 50,000 | 1,000 | $0.25 | $0.03 | 11% |
| Draft a long article Reads a little, writes a lot | 2,000 | 8,000 | $0.01 | $0.24 | 96% |
| Short answer after long reasoning 1,000 tokens of answer plus 20,000 of reasoning | 2,000 | 21,000 | $0.01 | $0.63 | 98% |
The summary costs $0.28 and nine tenths of that is input. The output ratio barely matters. The article costs about the same, $0.25, and almost all of it is output.
OpenAI's documentation on reasoning models (read on 14 November 2026) says the tokens a model spends thinking are not shown to you but "are billed as output tokens", and that a response can use "anywhere from a few hundred to tens of thousands" of them. So a short answer from a model that reasons first can cost more than a long article from one that does not. In our made-up example it comes to $0.64, and you only ever see 1,000 of the 21,000 tokens you paid for.
So if you use a reasoning mode, the output price is the one to check. The thinking is charged at the output rate, and the output rate has moved from three times the input rate to five.
Before you pick on price
- Look at the output price first if your work is writing, coding or anything that uses a reasoning mode. Look at the input price first if you mostly feed in long documents and ask short questions.
- Check what a "blended" price assumes. Sites that quote a single price per million tokens, ours included, have to assume a mix of input and output. Ours assumes one part input to three parts output and the formula is published. If your mix is different, the blended figure will mislead you.
- Measure a real job. Run your own task once, read your provider's usage report for that request, and only then scale it up. The cost calculator will do the multiplication.