Mistral Large 4 vs Claude Sonnet 5.5: The Promotion Is the Whole Price Argument

At list price this comparison is a rout. Mistral Large 4 costs $1.36 per million input tokens and $4.18 output against Claude Sonnet 5.5's $2.00 and $10.00, which is cheaper on both lines but not dramatically — 32% on input, 58% on output. With Mistral's launch promotion applied, the same three lines are $0.68, $0.07 and $2.09, and the gap becomes 66% on input and 79% on output. Same model, two very different arguments, and the difference between them is a date Mistral has not published.

The cards, as they read today

• Mistral Large 4 — $1.36 in, $0.14 cached, $4.18 out; promoted at half. 1.05 trillion parameters with 52 billion active, a 1.6-billion-parameter vision encoder, 1M context claimed and 524,288 measured. Text and image in, text out. Weights live under the Mistral Ecosystem Open-Weight License.

• Claude Sonnet 5.5 — $2.00 in, $10.00 out; 1M context; 128K maximum output; measured p50 first token 4.09 seconds. No published weights.

The output ceiling is the column that does not depend on a promotion. Mistral's measured 524K against Sonnet's 128K is a four-fold difference in how much a single request can produce, and for document generation that is the decider regardless of price.

What the promotion is worth per task

One million input tokens and 100,000 output tokens. At Sonnet's card: $2.00 of input plus $1.00 of output — $3.00. At Mistral's list: $1.36 plus $0.418 — $1.78. On promotion: $0.68 plus $0.209 — 89 cents.

So the same job costs $3.00, $1.78 or $0.89 depending on which vendor and which week. The promotion is doing more work in that sentence than the model does. Budget on $1.78 and treat $0.89 as upside; a plan built on the promotional number stops reconciling on a day nobody announces.

Where Sonnet's premium is not waste

Two things it buys that Mistral's card cannot match. First, latency: 4.09 seconds to first token is measurably fast for a flagship, and Mistral Large 4 has no comparable served measurement — a trillion-parameter MoE is not a low-latency product by design. Second, a stable, flat rate. A budget built on $2.00 and $10.00 today is still true next quarter; a budget built on a promo with no end date is a scenario, not a number.

What Sonnet does not buy is the licence. There is no checkpoint to run inside your perimeter, no derivative to train, no licence text to read — because there is no artefact. If that is the requirement, the price comparison is irrelevant and Mistral wins by existing.

How to choose

If you generate more than 128K tokens in a single request, the ceiling already decided. If you need ownership, the licence already decided. If neither applies, the decision is latency against price, and the only honest way to run it is on your own traffic: Sonnet 5.5 is on OrcaRouter at $2.00 and $10.00 with no markup, alongside 223 other models on one OpenAI-compatible endpoint; Mistral Large 4 is not in that catalogue and runs through Mistral's own API or third-party hosts.

And if you pick Mistral on price, put both numbers in the model — list for the plan, promotion for the optimism — because the one thing you can be certain of is that a promotion without an end date eventually has one.