Mistral Large 4 offers half-price API promotion with no end date

The pricing ambiguity

Mistral Large 4 has entered the market with a pricing structure that contains two different sets of numbers. The official list prices are $1.36 per million input tokens, $0.14 per million cached input, and $4.18 per million output. However, the pricing page also displays a launch promotion that strikes through these numbers to show a 50% discount.

Under this promotion, the costs drop to $0.68 for input, $0.07 for cached input, and $2.09 for output. As reported by OrcaRouter, there is no end date published for this offer anywhere on the site. This creates a situation where the current cost of an API call is roughly half the list rate, but nobody can say when that will change.

This ambiguity is significant because a 50% swing in a line item is often larger than the price difference between different models. For businesses planning their budgets, this means they must decide whether to build their financial models around the current discount or the higher list price. Planning for the day the promotion ends is a safer way to avoid being surprised by a sudden cost spike.

Understanding the costs

To understand what these numbers mean in practice, it helps to look at specific workloads. For a standard document pass involving one million input tokens and 100,000 output tokens, the list price is about $1.78. At the promotional rate, that same job costs roughly 89 cents, making the current promotion worth about 50% of the total cost.

Caching provides another way to manage these expenses. If a user reuses a large prefix of text, the cached rate is roughly a tenth of the standard input price. For a long agent run that re-reads its own history, using a cached prefix can reduce the input cost from $13.60 to about $1.40, which is a much larger saving than switching to a cheaper model.

Efficiency also plays a role in the final bill. Mistral reports that the model uses 47% fewer tokens per task than a comparable competitor on a coding agent index. Because users pay for tokens rather than for a model's rank on a leaderboard, this efficiency compounds with the current promotion to lower the total cost of a task.

Model architecture and limits

Mistral Large 4 is a sparse mixture-of-experts model. It features 1.05 trillion parameters in total, with 52 billion active parameters according to Mistral's own documentation. It also includes a 1.6-billion-parameter vision encoder, allowing it to process both text and image inputs to produce text outputs.

There is a discrepancy regarding the context window that users should notice. Mistral's card claims a 1M token context, but testing by Artificial Analysis shows a served window of 524,288 tokens. This distinction is important because a context window that cannot actually be filled is a context that will not be billed to the user.

Additionally, the public pricing page still contains a generic example for an older, smaller model. That example lists costs of $0.5 per million tokens in and $1.5 per million out. Anyone budgeting based on that marketing line instead of the specific Large 4 rate card will be off by a factor of three for both input and output.

Licensing and hosting

The model was released under the Mistral Ecosystem Open-Weight License, which is a modified MIT license. It allows for broad use, including commercial use and sublicensing, without a revenue threshold. However, it includes a specific restriction for anyone who operates a Model as a Service business.

A Model as a Service business is defined as giving a third party access to the model in a way that lets them exercise meaningful control over inputs or parameters. This means that while building a product that embeds the model's capabilities is allowed, standing up an API that hands out raw access to the model requires a separate commercial agreement. The cost of that agreement is not listed on the per-token price card.

For most teams, self-hosting this model is not a way to save money on tokens. The model weights are 1,060.6 GB in bf16, and running a trillion-parameter mixture-of-experts requires significant hardware. For almost any team below hyperscale, the cost of the GPUs and engineering time to host the model will likely exceed the cost of using the discounted API.

What to watch next

The most immediate thing to watch is the duration of the half-price window. Since there is no end date, the promotion could end without notice, making it a variable that needs to be monitored closely by those using the API. Budgeting for the list rate while treating the promotion as a bonus is the most defensible strategy for corporate spend.

Another factor to watch is how the 1M context window performs in practice. Because billed input follows the served window, it is important to see if the actual served limit remains consistent. If the served window is smaller than the advertised 1M tokens, it will affect how much data can be processed in a single call.

Finally, developers should watch how the license's 'meaningful control' clause is enforced in different scenarios. Because this phrase has no case law behind it, it remains an area of potential legal ambiguity. Companies should be careful to ensure their specific product structure does not fall into the category that requires a separate commercial license.