Mistral Large 4 vs DeepSeek V4 Pro: Two Open Flagships, and a Price Gap That Is Not the Point

On paper these two models are the same purchase: a trillion-parameter sparse mixture-of-experts, open weights, a European vendor on one side and a Chinese one on the other, and a per-token price that looks like the deciding column. It is not the deciding column. Mistral Large 4 and DeepSeek V4 Pro differ first on what they accept, second on what their licences let you build, and only third on cost — and the cost comparison itself is unstable, because one side is running a half-price promotion with no end date and the other doubles its rates for seven hours every weekday.

That reframing is the useful part, because the obvious comparison — $4.18 per million output against $1.98 — is true and misleading at the same time. Here is what each side actually is.

The specifications, side by side

• Size — Mistral Large 4 is 1.05 trillion parameters with 52 billion active by Mistral's own card, 49 billion by Artificial Analysis's reading, on a granular sparse MoE with 128 routed experts and four active per token. DeepSeek V4 Pro is 1.6 trillion with 49 billion active. The active counts land in the same place; the total counts do not.

• Modality — Mistral Large 4 is natively multimodal, with a 1.6-billion-parameter vision encoder, text and image in, text out. DeepSeek V4 Pro is text-only — on DeepSeek's own price sheet the vision row reads "Not supported" for Pro. This is the first real fork in the road, and it is not a matter of degree: a text-only model cannot do the job at all.

• Context and output — Mistral's card claims a 1M-token context; Artificial Analysis measured 524,288 tokens on the configuration it tests. DeepSeek V4 Pro lists 1M tokens of context and a 384K maximum output. Both figures are vendor ceilings on the Mistral side and a listed, served ceiling on the DeepSeek side.

• Licence — DeepSeek V4 Pro is MIT, with weights published and no field-of-use term. Mistral Large 4's weights landed on 29 October under the Mistral Ecosystem Open-Weight License, a 2,632-byte modified-MIT text that grants commercial use, modification and further training, but reserves one business model: clause 2 requires anyone operating a "Model as a Service" business to sign a separate commercial licence first. A gateway that merely relays requests to models hosted by others sits in a carve-out and is fine; a platform handing customers raw access to the model is not.

• Hosting footprint — Mistral Large 4 ships 1,060.6 GB in bf16 across 54 shards, with a 610.1 GB NVFP4 sibling that Mistral says runs on four B200 or B300 GPUs. DeepSeek V4 Pro is MIT-licensed at 1.6T parameters, which puts it in the same order of magnitude of hardware and a larger order of magnitude on memory.

• Independent scores — Mistral's card reports 26 on Artificial Analysis's Intelligence Index v5 and calls it the highest among western open-weight models. DeepSeek publishes no equivalent index figure of its own. We do not have a same-instrument, same-date reading for both models in hand, and a comparison built on two different instruments is exactly the mistake worth avoiding here — Mistral's own preview was reported at 38.4 on the pre-v5 index, so a naive before-and-after of its own numbers produces a twelve-point "regression" that does not exist.

The two price shapes are not the same kind of number

This is where the comparison stops being arithmetic and becomes scheduling.

Mistral Large 4 is priced flat: $1.36 per million input, $0.14 cached input, $4.18 output. It is currently running a launch promotion at exactly half those lines — $0.68, $0.07, $2.09 — with no published end date. The list price is stable and the effective price is not, and which one you pay depends on a date nobody has announced.

DeepSeek V4 Pro is priced by clock: $0.66 input and $1.98 output off-peak, doubling to $1.32 and $3.96 in two weekday windows, 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. The list price is the peak price twice a day and the discount the rest of the time.

Put next to each other at their best published rates, Mistral's promotional output price ($2.09) is almost identical to DeepSeek's off-peak output price ($1.98). At their worst, Mistral's list output ($4.18) is close to DeepSeek's peak output ($3.96) — but DeepSeek's peak happens for seven hours a day and Mistral's list price applies whenever the promotion ends. Neither number is the number.

One printed comparison makes the shapes concrete. Mistral's own API pricing page lists DeepSeek V4.1 Flash 1017 at $0.30 per million input and $1.20 per million output — exactly DeepSeek's peak rate, charged flat. The same model, on a reseller's card, is double DeepSeek's off-peak price and never cheaper than its peak one. That is not a knock on Mistral's pricing; it is what resale looks like when the upstream vendor moves its own rates twice a day.

What a task costs on each

One million input tokens and 100,000 output tokens, the shape of a document pass with a real answer at the end:

• Mistral Large 4 at list — $1.36 of input plus $0.418 of output: about $1.78.

• Mistral Large 4 on promotion — about 89 cents.

• DeepSeek V4 Pro off-peak — $0.66 plus $0.198: about 86 cents.

• DeepSeek V4 Pro in a peak window — about $1.72.

The same task, therefore, costs between 86 cents and $1.78 depending on which vendor and which hour — a spread of roughly 2× that has nothing to do with which model is better. If your workload can move in time but not in model, DeepSeek's schedule is a lever you control; if it cannot move in time, Mistral's promotion is a discount you receive and cannot plan on.

Where the decision actually turns

Three questions, in the order that resolves it.

Do you need images? If yes, the comparison is over: DeepSeek V4 Pro is text-only, and the Mistral side is the only side that accepts a picture. If no, this is the one dimension where the two are genuinely interchangeable, and the rest of the list decides.

Are you reselling raw access? If you are building a product whose features call the model, both licences are fine. If you are standing up an API that hands customers the model itself, Mistral's clause 2 requires a commercial agreement and DeepSeek's MIT does not. That is a legal difference with a price tag attached that neither vendor publishes.

Is your workload schedulable? If it is batch, a DeepSeek off-peak plan is arithmetic you can execute. If it is interactive and concentrated in one region's working hours, you are paying peak rates whether or not you intended to, and a flat-priced vendor is worth more than a cheaper one you cannot reach at the right hour.

The one thing that should not decide it is a leaderboard position. Mistral's 26 on Intelligence Index v5 is a vendor-reported reading of an instrument that rebuilt itself on 27 October; DeepSeek publishes no comparable figure. Two numbers from two instruments is not a comparison, it is a chart.

What you can call today

DeepSeek V4 Pro is on OrcaRouter at DeepSeek's own published rate — $0.66 per million input and $1.98 per million output off-peak, doubling exactly as DeepSeek's sheet defines it — through the same OpenAI-compatible endpoint and the same key as a 224-model catalogue, with automatic failover across provider paths. Mistral Large 4 is not in that catalogue as of publication, so this is not a route we can offer for it; it is callable through Mistral's own API and third-party hosts.

For a team choosing between them, the honest sequencing is: pick by modality and licence, then measure the token efficiency of each on your own traffic, then schedule. The price gap between the two is real but it is smaller at the rates you will actually be charged than at the rates on the two marketing pages — and the part that is genuinely different is what each licence lets you build afterwards.