- Artificial Analysis reports Mistral Large 4 costs $2.06 per Intelligence Index task, down 63% from the preview's $5.54 according to VentureBeat, and scores 26.
- The sources do not separate lower API rates from efficiency gains. A task's cost depends on both the rate and the number of tokens used.
- Before adopting the downloadable weights, check the license terms on model-as-a-service use and price the model on your own tasks.
A token price is a rate, not a bill
Most AI contracts are priced per token, a small chunk of text. But companies do not buy tokens. They buy finished tasks. The cost of a task is the rate multiplied by the number of tokens the model uses to finish it.
That means two things can move the bill: the price list and the model's appetite for words. Mistral's new Large 4 release gives a useful case. Independent tests now put numbers on both.
What was released
Mistral has published downloadable weights for Large 4 under a custom license, according to VentureBeat. Weights are the trained model itself. Whoever holds them can run the model on their own hardware.
The testing firm Artificial Analysis scored the model at its high reasoning setting. It reports 26 on its Intelligence Index v5.0, against a median of 23 for comparable open-weight models. The preview scored 25, VentureBeat reports. The weights arrived on October 29, two days after the planned date. VentureBeat says that came after about three more weeks of reinforcement learning and feedback from customers and security partners.
Cost per benchmark task fell 63%
The score moved one point. The cost moved far more. Artificial Analysis reports that running its Intelligence Index costs $2.06 per task on the new model. VentureBeat says the preview cost $5.54, a 63% drop.
That figure measures the benchmark's tasks. It is not the cost of a customer's own work.
VentureBeat also reports that Mistral's launch announcement introduces substantially lower API rates. So part of the fall may be a price cut. Artificial Analysis lists only current prices, so the sources do not show what the preview charged. They do not separate the price cut from any efficiency gain. They also do not say why cost fell.
Today's prices are $0.70 per million input tokens and $2.10 per million output tokens, according to Artificial Analysis. It calls both somewhat expensive against medians of $0.47 and $1.95. It also calls the model somewhat verbose: 110 million tokens on the index, against a median of 99 million.
Token use is the second lever
VentureBeat points to a sign that token use matters here. In Artificial Analysis's coding-agent test, Large 4 is second among current open-weight entries. The GLM-5.3 setup that leads it consumes about twice as many tokens.
The lesson here is that a buyer should ask for two numbers, not one: the rate and the tokens a finished job consumes. A rate cut and a leaner model both show up as a cheaper task, and only one of them is a lasting trait of the model.
How the model is built
Large 4 is a Mixture of Experts model. Artificial Analysis lists about 1.1 trillion total parameters, the adjustable numbers a model learns in training. VentureBeat cites 1.05 trillion from the firm's release assessment. Only 52 billion are active for any given request.
A routing step picks a small group of specialist sections for each token. The rest sit idle. That is how a very large model can still respond quickly. Artificial Analysis measures 161 tokens per second on Mistral's API against a median of 73, and 1.42 seconds to the first token against 1.60.
Ownership has a license attached
Mistral's chief revenue officer, Marjorie Janiewicz, told VentureBeat what customers ask for. "I don't want to get disconnected. I want to own my model," she said. She added that customers want to customize it and that they care about business continuity.
That is a risk argument. A hosted model can change price or terms, or be withdrawn. Downloaded weights remove that dependency.
The license still matters. VentureBeat reports that the Mistral Ecosystem Open-Weight License has no revenue threshold. A separate commercial agreement is required if the user or an affiliate runs a model-as-a-service business, with specific exceptions. Artificial Analysis summarizes it as commercial use allowed with restrictions. Companies that resell AI features to customers should have counsel read it first.
Where the lead stops
The headline claim has limits. VentureBeat reports that the model scores one point above Muse Spark Open at its highest setting. That makes it the top-scoring open-weight model developed outside China in that evaluation.
DeepSeek V4.1 Flash 1017 and GLM-5.3 sit six points ahead, each run at its maximum reasoning setting. VentureBeat cautions that results can shift with the software and settings used to run the model. Artificial Analysis lists only one API provider, so there is little price competition yet.
What to ask your team
First, ask what a completed task costs on your own workload, not the benchmark's. Run a sample of real jobs and count both the rate and the tokens.
Second, ask which provider dependency you would actually remove by self-hosting, and what the hardware and staff cost to do it. Third, ask whether your business model triggers the license's model-as-a-service clause.
Last, repeat the test whenever a vendor ships a new version or changes its rates. Between preview and release, benchmark cost fell 63% while the score rose one point. The price of a token tells you little. The cost of the task tells you what you are paying.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Artificial Analysis.
