Overview of the prices of the main large language models depending on the number of tokens requested via their API. Prices can vary completely depending on the offers.
LLMs are billed by “token”, a unit of measurement representing segments of text, that is, words or fragments of words. In general, large language models (LLM) price input tokens (those sent to the model) and output tokens (those generated in response) separately. This billing method allows users to estimate the cost based on the volume of data processed by the LLMs made available in the form of an API. According to OpenAI's definition, a token corresponds, in English, to 0.75 words. The table below summarizes the estimated prices per token for the main LLMs on the market, whether entry or exit.
| Model | Supplier | Price for 1000 input tokens | Price for 1000 tokens output | Total number of tokens supported |
|---|---|---|---|---|
| GPT-4o (omni) | OpenAI | 0,05 dollar | 0,15 dollar | 128 000 |
| GPT-4 Turbo | OpenAI | 0,1 dollar | 0,3 dollar | 128 000 |
| GPT-4 | OpenAI | 0,3 dollar | 0,6 dollar | 8 000 |
| Claude 3 Haiku | Anthropic | 0,0025 dollar | 0,0125 dollar | 200 000 |
| Claude 3 Sonnet | Anthropic | 0,03 dollar | 0,15 dollar | 200 000 |
| Close 3 Work | Anthropic | 0,15 dollar | 0,75 dollar | 200 000 |
| Call 3 70b | Meta (via AWS) | 0,00265 dollar | 0,0035 dollar | 8 000 |
| Call 2 70b | Meta (via AWS) | 0,00195 dollar | 0,00256 dollar | 4 000 |
| Gemini 1.0 Pro | 0,005 dollar | 0,015 dollar | 32 000 | |
| Gemini 1.5 Pro | 0,07 dollar | 0.21 dollar | 1 000 000 | |
| Command | Cohere | 0,1 dollar | 0,2 dollars | 4 000 |
| Command R | Cohere | 0,005 dollars | 0,015 dollars | 132 000 |
| Command R+ | Cohere | 0,03 dollars | 0,15 dollars | 128 000 |
| Mixtral 8x7B | Mistral AI (via Anyscale) | 0,005 dollars | 0,005 dollars | 32 000 |
| Mistral Small | Mistral AI | 0,02 dollars | 0,06 dollars | 32 000 |
| Mistral Large | Mistral AI | 0,08 dollars | 0,24 dollars | 32 000 |
| GPT-3.5 Turbo | OpenAI | 0,12 dollars | 0,16 dollars | 4 000 |
| PaLM 2 | 0,02 dollars | 0,02 dollars | 8 000 |
Model Analysis
- GPT-4o et GPT-4 Turbo (OpenAI) – Within OpenAI's LLM range, GPT-4o is a high-performance multimodal version at a competitive price, while GPT-4 Turbo is a lightweight option at reduced costs. GPT-4 remains the main version for complex tasks, but at a higher cost.
- Gemini 1.5 et 1.5 Pro (Google DeepMind) – Developed by Google, Gemini templates are optimized for text and image. The Pro version is more powerful, with higher token capacity for advanced use cases.
- Claude 3.5 Sonnet, Haiku et Opus (Anthropic) – Anthropic offers different versions of its LLM for specific needs, rapid responses (Haiku) or long and in-depth interactions (Sonnet and Opus). Prices vary depending on the abilities of each person.
- Call 3 70b and Call 2 70b (Goal) – Developed by Meta, these models are notably offered by the Amazon cloud. They provide increased flexibility for custom tasks.
- Command, Command R, et Command R+ (Cohere) – Cohere offers a full range of models for research and analysis, ranging from cost-effective options for basic data research to LLMs tailored for more complex analyses.
- Mixtral and Mistral – Mistral French models are optimized for specific linguistic tasks (Mixtral) or quick interactions (Mistral Small).
- GPT-3.5 Turbo – This is a cost-effective option of OpenAI models, suitable for simple tasks with limited token capacity, offering a good alternative for less intensive needs.
