Rent It or Own It? — Mistral Large 4 — Against Claude Opus
Mistral Large 4 is a trillion-parameter model you can rent today and, Mistral says, download by the end of October. It costs less per token than Medium 3.5. Rent it, keep paying for Claude Opus, or wait for the weights? One number decides it, and it's in last month's usage report.
The Promise
- The preview API costs $1.36 per million input tokens and $4.18 per million output. On output that's 44% less than Mistral Medium 3.5, a model nearly eight times smaller.
- On a typical agent workload (80% of tokens in, 20% out), a million tokens costs about $1.92 on Large 4 and $7.20 on Claude Opus 5.5.
- Mistral trained Large 4 in its own European data centers, serves the preview from them, and says it will run a deployment end to end under European law.
- Weights on your own servers don't change their terms overnight. Mistral reports that Claude Opus 5.5 and GPT-6 Astra refused its vulnerability-reproduction test; that's Mistral's test and framing, but a refusal mid-incident is a real failure mode.
The Risk
- One trillion parameters is about 500 GB at 4-bit before any context. No consumer GPU comes close, and a 512 GB Mac Studio runs out of room.
- Owning means four NVIDIA B200s at about $22,500 a month. You'd need roughly 11.7 billion tokens a month to beat Mistral's own API, and about 3.1 billion to beat Claude Opus 5.5.
- Large 4's license hasn't been published. Until it is, open weights is a promise, and you can't put a promise in a contract.
- The blind-test and cyber figures are Mistral-reported. In Mistral's own blind test, Claude Opus 5 still scored 4.22 out of five against Large 4's 3.74.
Rent, keep paying, or wait
On 6 October, Mistral shipped a trillion-parameter model. You can rent it now and, if Mistral keeps its word, own the weights by the end of October. It costs less per token than Mistral Medium 3.5, which I covered in May. In a blind coding test Mistral commissioned, only Claude Opus scored higher.
So the decision this month has three options: rent Large 4, keep paying for Claude, or wait three weeks for the weights. I ran the arithmetic on all three.
Does it fit?
In May, the answer for Medium 3.5 was yes: 128 billion parameters, about 64 GB at 4-bit, and one Mac Studio held it. Large 4 is a mixture-of-experts model with one trillion parameters in total and 49 billion active per token, nearly eight times the size.
The two numbers do different jobs. The active count decides speed. The total decides memory, because every expert has to sit in memory while it waits its turn. Mistral says the weights ship in 8-bit and 4-bit, and at 4-bit a trillion parameters is about 500 GB before a single token of context. No consumer GPU comes close, and a 512 GB Mac runs out of room. A community 3-bit build would be roughly 375 GB, which a 512 GB Mac Studio could hold with some quality lost. How much is unknown until the weights are out.
For almost everyone, Large 4 is a model you rent.
What renting costs
Mistral’s preview API charges $1.36 per million input tokens and $4.18 per million output. Medium 3.5 costs $1.50 in and $7.50 out, and price was the complaint in May. The model nearly eight times bigger now costs 44% less on output. It’s a preview price, and Mistral can change it.
Claude Opus 5 topped Mistral’s blind test; its successor, Claude Opus 5.5, rents from Anthropic at $4 in and $20 out. Assume an agent workload with 80% of tokens going in and 20% coming out (my assumption, so check it against your logs). A million tokens costs about $1.92 on Large 4 and $7.20 on Claude Opus 5.5, 3.7 times more. The two use different tokenizers, so treat that as a ratio.
What owning costs
Mistral says the 4-bit weights are meant to run on four to eight NVIDIA B200s. Each carries about 180 GB, so four hold the 500 GB model with room for context. On 4 October, the Ornn Index, built from actual rental deals, put a B200 at $7.71 an hour. Four of them around the clock come to about $22,500 a month.
Divide by the rental price. To beat Mistral’s own API, you’d need to push about 11.7 billion tokens a month through those four GPUs. To beat Claude Opus 5.5, about 3.1 billion. Whether four B200s can serve that much can only be measured once the weights ship. Below a few billion tokens a month, owning the weights won’t save money. It buys something else.
What owning buys
Where your data goes. Mistral trained Large 4 in its own European data centers and serves the preview from the same machines, and says it will run a European deployment end to end under European law. If the US CLOUD Act reaching your prompts is the risk your lawyers keep raising, that matters. It doesn’t free you from NVIDIA: Large 4 was trained on 3,800 Grace Blackwell GPUs, and the self-hosting guidance is NVIDIA hardware.
Access you control. Mistral ran a test that asks a model to reproduce a real vulnerability in open-source software, then patch it, and reports that Claude Opus 5.5 and GPT-6 Astra scored near zero because they refused. That’s Mistral’s test and Mistral’s framing. In my experience, though, the worst moment to discover a vendor’s policy has changed is the middle of an incident. Weights on your own servers don’t change their terms overnight.
The license is the part you can’t count on yet. Medium 3.5 shipped under a modified MIT license; Large 4’s hasn’t been published. The preview API is also monitored, and Mistral blocks requests it reads as malicious cyber work. Vetted security teams and state authorities are testing a version with reduced moderation. Everyone else waits for the weights.
The verdict
European organizations with data-residency rules should put Large 4 on the short list and test the preview this month, as should security teams that can’t afford a refusal halfway through an incident. Developers happy on Claude should weigh the price gap against the quality gap Mistral itself reported: 4.22 out of five for Claude Opus 5 against 3.74 for Large 4.
In May the internet shrugged at Mistral, and I said: “That shrug should worry you.” The complaint then was price, and Large 4 settles it. Whether it becomes the model you own depends on a license nobody has read yet. Rent it this month; decide on owning when the weights and the license arrive.
The question for your next budget review: how many tokens did we run last month? Under three billion, you’re renting either way. The only question left is from whom.