What Does Mistral Small 4 Cost?
Mistral Small 4 launched on March 16, 2026 at $0.15 per million input tokens and $0.60 per million output tokens on Mistral's API. The weights ship under Apache 2.0, so self-hosting carries no licence fee at all.
That combination is the whole story. A model this cheap on the API is usually closed, and a model this open usually has no first-party endpoint worth using. Small 4 has both.
Here is what Mistral published at launch:
| Attribute | Mistral Small 4 |
|---|---|
| Launch date | March 16, 2026 |
| Total parameters | 119B |
| Context window | 256K tokens |
| Licence | Apache 2.0 |
| Modality | Multimodal, via the Pixtral capability folded into the weights |
| API price, input | $0.15 per million tokens |
| API price, output | $0.60 per million tokens |
Every figure above comes from Mistral's own launch post. AI Perks tracks what the same providers give away in credits on top of list price.
What Merging Magistral, Pixtral and Devstral Actually Changes
Small 4 is the first Mistral model to fold reasoning, multimodal and agentic coding into one set of weights. Mistral describes it as unifying Magistral for reasoning, Pixtral for multimodal and Devstral for agentic coding.
The practical effect is on your architecture, not your benchmark spreadsheet. Before this, a team doing document parsing plus code generation plus reasoning ran three model families, three sets of prompts and three billing lines.
Three things collapse into one:
One endpoint for mixed workloads. The multimodal capability that used to sit in Pixtral now lives in the same weights, so mixed media work no longer needs a second model behind a second endpoint.
One set of weights to host. If you self-host, you were previously pulling and serving separate checkpoints. Now it is one.
One price to model. Cost forecasting across a mixed pipeline stops requiring a weighted blend of three rate cards.
The 256K context window is what makes the consolidation usable. Feeding a long document and a codebase into the same request is only interesting if the window holds both.
Mistral Small 4 Pricing: What Real Workloads Cost
At $0.15 input and $0.60 output, a serious production workload lands in the tens of dollars per month, not the thousands. The arithmetic below is list price multiplied by volume, nothing more.
| Monthly workload | Input tokens | Output tokens | Cost at list price |
|---|---|---|---|
| Prototype or side project | 5M | 1M | $1.35 |
| Support chatbot | 50M | 10M | $13.50 |
| Document parsing pipeline | 300M | 15M | $54.00 |
| Coding agent, one developer | 500M | 100M | $135.00 |
| Production RAG at scale | 2B | 200M | $420.00 |
The output multiplier is the number to watch. Output costs 4x input, so a chatty agent that narrates its reasoning costs multiples of a terse one doing identical work.
That ratio is also why long context is cheaper than it sounds. Stuffing 200K tokens of documents into a prompt costs $0.03. It is the response that bills.
Credits change this math, and the free routes in the next section cover the first few million tokens outright.
The Free Routes to Mistral Small 4
There is no dedicated free API tier for Mistral Small 4. There are three separate free or near free routes to running it, and they suit different stages.
| Route | What it costs | What you get |
|---|---|---|
| NVIDIA developer prototyping | Free | Free prototyping on NVIDIA accelerated computing, named in Mistral's launch post |
| Mistral Free consumer plan | $0 | $10 per month in API credits |
| Self-hosted open weights | Your own compute | Apache 2.0, no per-token fee |
| Provider credit programs | Free to hold | Tracked at AI Perks |
Mistral has not published paid plan credit allowances beyond that $10 figure, so any tier by tier credit table you find elsewhere is someone's reconstruction rather than a vendor number.
The $10 per month credit on the free consumer plan is the underrated one. Run it through the list price above and it buys roughly 66 million input tokens if you spend it purely on input. Assume a 5:1 read to write ratio and it lands nearer 44 million tokens. Both are arithmetic on the published rate, not a published token allowance, but for an evaluation harness or a low volume internal tool it is a real budget rather than a token gesture.
One caveat worth stating plainly: these are consumer plan credits, and Mistral has not published guarantees about how the allowance behaves under sustained load. Treat it as evaluation budget, not production capacity.
Apache 2.0: What Self-Hosting Actually Changes
Apache 2.0 means you can run, modify, fine-tune and commercially deploy Mistral Small 4 with no licence fee and no usage reporting. It does not mean running it is free.
The bill moves rather than disappearing. Open weights run on whatever inference stack you already operate, and every one of those stacks runs on hardware someone pays for. Mistral has not published a deployment cost model for Small 4, so the comparison below is about your own duty cycle rather than a vendor figure.
The break-even is set by duty cycle, not by list price:
- Bursty or low volume traffic favours the API, because idle GPUs bill at full rate while an idle API call bills nothing.
- Sustained high volume traffic favours self-hosting, because per-token pricing scales linearly while a reserved GPU does not.
- Data residency or air-gapped requirements decide it outright, and price is not the variable.
The genuinely useful part of an Apache 2.0 release is not saving $0.15 per million tokens. It is that nobody can deprecate the model out from under you. A closed model you built a product on can be retired on the vendor's schedule. This one cannot.
Compute credits are what make self-hosting cheap, and they are a separate grant from model credits. AI Perks covers both layers.
What Mistral Has Not Published
Three things about Small 4 and its siblings are commonly stated online with more confidence than the vendor has earned. Here is what is actually unconfirmed.
Startup credits. Mistral has run a startup credit program, reported historically at up to around $30,000 in La Plateforme credits. That figure is reported rather than currently confirmed: as of mid 2026 there is no public self-serve page for the program, so treat any specific number you see quoted as unverified until Mistral republishes one. AI Perks tracks which model providers currently have live credit programs.
Mistral OCR 4. A document intelligence model dated June 23, 2026 is reported but not confirmed on a versioned pricing page, so read the model and its date as reported rather than settled. Mistral's general pricing approach for OCR is billing per 1,000 pages, and the exact per 1,000 page rate for this version is not confirmed. Anyone quoting a precise number for it is guessing.
Free tier limits. There is no dedicated free API tier for Small 4. The $10 per month allowance sits on the consumer Free plan, and Mistral has not published rate limits for it, so there is no published ceiling to design against.
If a page states a figure the vendor never published, it will be wrong the moment the vendor publishes one.
Frequently Asked Questions
How much does Mistral Small 4 cost per million tokens?
Mistral Small 4 costs $0.15 per million input tokens and $0.60 per million output tokens on the Mistral API. Output bills at 4x input, so verbose agents cost considerably more than terse ones for identical work. Self-hosting the Apache 2.0 weights carries no per-token fee, only compute cost.
Is there a Mistral AI API free tier in 2026?
There is no dedicated free API tier for Mistral Small 4. The consumer Free plan includes $10 per month in API credits. Mistral has not published rate limits for that allowance, and it has not published credit allowances for its paid consumer tiers either, so quoted per tier numbers are unverified.
Can I run Mistral Small 4 for free?
Yes, through three routes. Mistral's launch post points developers to free prototyping on NVIDIA accelerated computing. The consumer Free plan carries $10 per month in API credits. And the Apache 2.0 weights can be self-hosted on your own hardware at no licence cost.
What models did Mistral Small 4 replace?
Mistral describes Small 4 as the first model unifying Magistral for reasoning, Pixtral for multimodal and Devstral for agentic coding into one model. It is multimodal through the folded in Pixtral capability, has a 256K context window, and runs 119B total parameters.
Does Mistral offer startup program credits?
Mistral has run a startup credit program reported historically at up to around $30,000 in La Plateforme credits. That is a reported figure, not a currently confirmed one: there is no public self-serve page for the program as of mid 2026, so treat any quoted amount as unverified. AI Perks tracks which providers have live programs and what they are currently worth.
Is self-hosting Mistral Small 4 cheaper than the API?
It depends on duty cycle, not list price. Bursty or low volume traffic favours the API, since idle GPUs bill while idle API calls do not. Sustained high volume favours self-hosting. Compute credits shift that break-even sharply, and they are a separate grant from model credits.
The cheapest token is the one someone else already paid for. Find them at AI Perks.