# Why do output tokens cost more than input tokens on LLM APIs?

Published 2026-11-17. Prices as of 2026-11-22.

Output tokens cost more because a model writes them one at a time, while it reads input tokens in parallel. As of Nov 22, 2026, the median text model in Sovyron's catalog charges 5× as much per output token as per input token on the maker's own API. A request that mostly reads costs less than one that mostly writes.

## Why do providers charge more for output tokens?

A model handles the two kinds of tokens in different ways. NVIDIA describes the prefill phase, where the model processes the whole prompt, as a highly parallelized matrix-matrix operation that saturates GPU utilization. In the decode phase the model generates output tokens autoregressively, one at a time, and each token needs the keys and values of all the tokens before it. NVIDIA calls that a memory-bound operation that underuses GPU compute compared with prefill.

Prompt tokens can be processed together, while each output token has to wait for the one before it. That is the usual technical explanation for the price gap. No provider publishes how it sets its rates, so read it as an explanation of the pattern and not as a pricing formula. Providers list the two rates in separate columns, for example on [Anthropic's pricing page](https://platform.claude.com/docs/en/about-claude/pricing).

## How much more do output tokens cost?

Sovyron compares the two rates on the maker's own API for 141 text models, leaving out speech, realtime, image-generation and omni models. The median output price is 5× the input price. 39 models sit at exactly 5×, and 124 of the 141 fall between 3× and 8×. The widest gap is 8.3×, for example on Gemini 2.5 Flash. 5 models charge the same price both ways, for example Ministral 3 14B.

| Output ÷ input | Models | Share of models |
| --- | --- | --- |
| 1× (same price) | 5 | 4% |
| Above 1× to under 4× | 37 | 26% |
| 4× to under 5× | 20 | 14% |
| 5× to under 6× | 43 | 30% |
| 6× to 8× | 33 | 23% |
| Above 8× | 3 | 2% |

| Model | Input / 1M | Output / 1M | Output ÷ input | Maker |
| --- | --- | --- | --- | --- |
| [GPT-6.2 Astra](https://sovyron.com/models/gpt-6-2-astra/) | $8 | $40 | 5× | OpenAI |
| [Qwen3.8 Flash](https://sovyron.com/models/qwen3-8-flash/) | $0.15 | $0.47 | 3.1× | Alibaba |
| [Gemini 4 Neon](https://sovyron.com/models/gemini-4-neon/) | $3 | $15 | 5× | Google |
| [Mistral Large 4](https://sovyron.com/models/mistral-large-4/) | $0.7 | $2.1 | 3× | Mistral |
| [GLM-5.3-FlashX](https://sovyron.com/models/glm-5-3-flashx/) | $0.37 | $1.25 | 3.4× | Z.AI |
| [Claude Fable 5.2](https://sovyron.com/models/claude-fable-5-2/) | $8 | $40 | 5× | Anthropic |
| [Muse Ember](https://sovyron.com/models/muse-ember/) | $0.1 | $0.5 | 5× | Meta |

## What the output price does to a real bill

The split between input and output decides which rate drives the bill. Take Claude Sonnet 5.5 on Anthropic: input costs $2 per 1M tokens and output costs $10. The table prices 1M tokens at three different splits. The splits are illustrations, so replace them with your own traffic.

| Workload | Input / output split | Cost per 1M tokens | Share of cost from output |
| --- | --- | --- | --- |
| Summarizing long documents | 95% / 5% | $2.40 | 21% |
| Chat | 70% / 30% | $4.40 | 68% |
| Drafting from short briefs | 20% / 80% | $8.40 | 95% |

A workload that reads a lot and writes a little is cheap per token. One that writes a lot costs several times more for the same token count. The blended rate on each model page assumes three input tokens for every output token, which the [methodology page](https://sovyron.com/methodology/#blend) explains. If your traffic looks different, price it with your own split.

## Do reasoning tokens count as output tokens?

Yes. OpenAI says reasoning tokens are not visible through the API but occupy space in the context window and are billed as output tokens. Anthropic's extended thinking guide says the response usage field reports how many of the billed output tokens were internal reasoning.

Here is what that does on Claude Sonnet 5.5. A request sends 2,000 input tokens and gets 500 tokens of visible answer. Without reasoning it costs $0.0090. If the model also spends 3,000 tokens thinking, the billed output is 3,500 tokens and the request costs $0.0390, about 4.3 times as much. The 3,000 thinking tokens are an assumption for the example. Real counts vary by prompt and by model.

## How to estimate and lower the output share of your bill

1. Estimate input tokens. OpenAI gives these rules of thumb for English text: 1 token is about 4 characters, and 100 tokens are about 75 words. Other languages differ. For exact counts, use the provider's token counting API.
2. Estimate output tokens separately, and add the reasoning tokens if the model thinks before it answers. Check a few real responses for the usage numbers instead of guessing.
3. Multiply each by its own rate and add them: (input tokens × input price + output tokens × output price) ÷ 1,000,000. The calculator on each model page, such as the [Claude Sonnet 5.5 page](https://sovyron.com/models/claude-sonnet-5-5/), does the same sum with your volumes.
4. Cap the output. OpenAI's max_output_tokens limits all generated tokens, reasoning included, and Anthropic's Messages API has max_tokens. Ask for short formats such as a label, a score or a few bullets when that is all you need.
5. Compare models on the output price too, and look past the input price alone. The [cheapest LLM API ranking](https://sovyron.com/cheapest/llm-api/) uses the blended rate, and [why the same LLM costs different amounts](https://sovyron.com/articles/same-llm-different-price/) covers price differences between providers. If long prompts repeat, [prompt caching](https://sovyron.com/articles/prompt-caching-savings/) lowers the input side.

## FAQ

### Why are output tokens more expensive than input tokens?

The model reads the prompt in one parallel pass, then writes the answer one token at a time, with each token depending on the ones before it. Generation uses the hardware less efficiently. Providers do not publish their pricing formulas, so this is the standard technical explanation of the gap.

### How much more do output tokens cost than input tokens?

Across 141 text models in Sovyron's catalog, the median output price is 5× the input price on the maker's own API, as of Nov 22, 2026. 88% of them fall between 3× and 8×. Check the exact pair for a model on its [model page](https://sovyron.com/models/claude-sonnet-5-5/).

### Do all LLM APIs charge more for output?

Nearly all do. 5 of the 141 text models in the catalog charge the same price for input and output, for example Ministral 3 14B. The rest charge more for output.

### Are thinking tokens charged as output tokens?

OpenAI states that reasoning tokens are billed as output tokens, and Anthropic reports thinking tokens as part of the billed output tokens in the usage field. A model that thinks a lot can cost several times more per request than the visible answer suggests.

### How can I cut the cost of output tokens?

Ask for shorter answers, set a cap such as max_output_tokens, lower the reasoning effort where the provider lets you, or pick a model with a cheaper output rate. The [cheapest LLM API ranking](https://sovyron.com/cheapest/llm-api/) is a starting point for the last option.

## Sources

- [NVIDIA: Mastering LLM Techniques, Inference Optimization](https://developer.nvidia.com/blog/mastering-llm-techniques-inference-optimization/)
- [OpenAI reasoning models guide](https://developers.openai.com/api/docs/guides/reasoning)
- [Anthropic extended thinking docs](https://platform.claude.com/docs/en/build-with-claude/extended-thinking)
- [Anthropic pricing page](https://platform.claude.com/docs/en/about-claude/pricing)
- [OpenAI: What are tokens and how to count them](https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them)

HTML page: https://sovyron.com/articles/output-tokens-cost-more/
Methodology: https://sovyron.com/methodology/#blend
