Mistral Large 4 vs Gemini 3.8 Flash: The Promotion Makes It Cheaper Than Google's Cheap Tier

On paper this is a flagship against a flash model and should not be close. In practice, right now, Mistral Large 4's promotional rate — $0.68 per million input and $2.09 output — undercuts Gemini 3.8 Flash's $0.75 and $3.75 on both lines, while offering a larger output ceiling. At Mistral's list price of $1.36 and $4.18 the ordering flips on input and nearly flips on output. So the honest question is not which model is cheaper. It is which price you will be paying in three months, and nobody at Mistral has published the date that decides it.

The three cards that matter

• Mistral Large 4 at list — $1.36 in, $0.14 cached, $4.18 out per million. 1M context claimed, 524,288 measured. Text and image in, text out. Weights published on 29 October, 1,060.6 GB bf16, under the Mistral Ecosystem Open-Weight License.

• Mistral Large 4 on promotion — $0.68, $0.07, $2.09. Same model, half the price, no end date published anywhere we can find.

• Gemini 3.8 Flash — $0.75 in and $3.75 out per million, with cache reads at $0.075, on OrcaRouter at Google's list price. 1M context, but a 65K maximum output, and a measured p50 time to first token of 8.13 seconds.

Two of those three are the same product at different prices, which is why this comparison has to be run twice.

The output ceiling is the real difference

Sixty-five thousand tokens sounds like a lot until you try to generate a long document in one call. A 40,000-word report is roughly 55,000 tokens; add an outline and a preamble and a single-call generation on Gemini 3.8 Flash is at its ceiling before it has finished. Mistral's measured 524K window is eight times larger, and its vendor-claimed 1M is more than that again.

For chat and agent traffic, the ceiling never binds and both models behave the same. For document generation, transcript synthesis, long code files or anything that asks one request to produce the whole artefact, the ceiling is the deciding column and the price is a rounding error on top of it. That is the honest framing of "flash versus flagship" here: the small model is not cheaper for the work it cannot do.

Latency, and whose number to trust

Gemini 3.8 Flash measures 8.13 seconds to first token on our page. Mistral Large 4 has no comparable served measurement we can cite — Artificial Analysis reports 524,288 tokens of context on its tested configuration, but a trillion-parameter MoE is not a low-latency product in the way a flash tier is designed to be. If your product is interactive, assume the larger model is slower and measure before committing; if it is batch, neither number matters.

Where the licence enters, and why it is not a footnote

Mistral's weights are downloadable under a licence that permits commercial use, modification and further training, and reserves model-as-a-service — giving third parties meaningful control over inputs, parameters or training data — for a separate agreement. Google's model has no such document because there is nothing to download. If the reason you are looking at Mistral is that you want to run the model inside your own perimeter, that is the whole difference and the price comparison is secondary.

How to decide

If you need a large output ceiling this quarter and you are willing to accept a promotion that may end, Mistral Large 4 at $0.68/$2.09 is the cheapest large-context multimodal generation available on these two cards — and the risk is that the price doubles without notice, which is exactly the thing to write into a budget as a scenario rather than a surprise.

If you need a flat, published, cache-discounted rate with no licence to reason about, Gemini 3.8 Flash is $0.75 / $3.75 with cache reads at $0.075, on OrcaRouter at Google's list price with no markup, on the same key as 223 other models. It is also the cheaper option the moment Mistral's promotion ends, and the cheaper one today for any workload that never approaches its 65K ceiling only in the sense that it is not paying for a capability it cannot use.

Run both for a week on the same prompts and count completed tasks per dollar. The sticker comparison has already been made twice in this article, and it gave two different answers.