Mistral Large 4 vs Grok 4.7: Same Context Class, Opposite Ownership

Put the two cards on one screen and the price difference is the first thing you see and the last thing that matters. Mistral Large 4 lists at $1.36 per million input and $4.18 output — currently half, with no end date — and Grok 4.7 sits at $2.00 and $6.00 with a 500K-token context and a 450K output ceiling. Output is 30% cheaper on Mistral at list and 65% cheaper on promotion; input is cheaper still. But one of these models can be downloaded and the other cannot, and that is the column that survives a price change.

The two cards, without the marketing

• Mistral Large 4 — 1.05 trillion parameters, 52 billion active by Mistral's card; a 1.6-billion-parameter vision encoder; 1M context claimed and 524,288 measured; $1.36 in, $0.14 cached, $4.18 out, promoted at half. Weights live on Hugging Face, 1,060.6 GB bf16, ungated, under the Mistral Ecosystem Open-Weight License.

• Grok 4.7 — 500K context, 450K maximum output, $2.00 in and $6.00 out per million, served by SpaceXAI through its own API. No published weights, no published parameter count, no open repository.

The context numbers look far apart until you apply the same scepticism to both. 500K is what Grok states and 1M is what Mistral states; the independent measurement of Mistral Large 4 found 524,288 tokens on the tested configuration, which is within 5% of Grok's ceiling. In practice these two models offer roughly the same working window, and the 1M headline is a vendor figure rather than a served one.

Why the price gap is smaller than it reads

Two corrections make the comparison honest. The first is the promotion: at $4.18 against $6.00, Mistral is 30% cheaper on output, not 65%. The second is token efficiency. Mistral reports 47% fewer tokens per task than a comparable competitor on Artificial Analysis's coding agent index; if even half of that holds against Grok on your traffic, the per-task gap widens well beyond the sticker difference, and if none of it holds, the sticker difference is all you get.

Grok's card has one structural advantage Mistral's does not: it is flat. There is no weekday window in which the rate doubles and no promotion that can expire. A budget built on $2.00 and $6.00 stays true next month. A budget built on $0.68 and $2.09 has a date-shaped hole in it that nobody at Mistral has filled.

The ownership question, which is the real one

Mistral Large 4's licence is not Apache and not MIT. It is the Mistral Ecosystem Open-Weight License: a 2,632-byte modified-MIT text that grants commercial use, modification, distribution, sublicensing and further training, and reserves exactly one business model — operating a model-as-a-service where third parties get meaningful control over inputs, parameters or training data requires a separate commercial agreement. Embedded features, internal use and mere relaying are carved out.

Grok 4.7 offers none of that. There is no checkpoint to run in your own perimeter, no derivative you may train, and no licence to read because there is no artefact. If your requirement is data residency or the right to fine-tune, the comparison ends at that sentence, regardless of price.

If your requirement is simply to call a large model with a long context, then you are comparing two flat-ish prices on two models whose quality claims come from different instruments — Mistral's own release post and Artificial Analysis for the coding index, and SpaceXAI's own numbers for Grok — and neither vendor has published a head-to-head on a common task set. That is the normal state of this market, and it means the only reliable comparison is the one you run.

What to do with it

Grok 4.7 is on OrcaRouter at SpaceXAI's list price of $2.00 and $6.00, alongside 223 other models on one OpenAI-compatible endpoint with automatic failover. Mistral Large 4 is not in that catalogue, so this is not a route we can offer; it is callable through Mistral's own API and third-party hosts.

The decision rule that survives both cards: if you need ownership, choose Mistral and budget at list price in case the promotion ends. If you need a flat rate and a long context with no licence to reason about, choose Grok. If you need to know which is better on your tasks, route a slice of traffic at each and count output tokens per completed job — the one number neither vendor publishes for the other's model.