Mistral Large 4: API Pricing, Tasks and Hosting Costs

ZenMux’s October 8, 2026 directory lists model ID mistralai/mistral-large-4 at $0.68 per million input tokens and $2.09 per million output tokens, marked 50% OFF. The directory gives no end date for that offer. Treat this as the current discounted channel quote rather than a permanent tariff.

The listed channel context is 256K, with a displayed maximum output of 256K. A maximum-output label does not mean a full prompt and a 256K answer both fit at once: leave room inside the route’s context. Mistral’s official 1M window describes its direct route, not a 1M entitlement on ZenMux.

Use the exact ZenMux ID in the gateway’s request format and check that route’s accepted limits before sending a large document. The directory shows one provider. No ZenMux cache rate is quoted here; Mistral’s direct cached-input rate should not be applied to this gateway bill.

Mistral lists mistral-large-4 with a 1M-token context. OpenRouter uses mistralai/mistral-large-4-0 and lists a 524,288-token context with up to 262,144 output tokens. Size documents and reserved output against the route you actually call.

Mistral Large 4 accepts text and images and generates text. Its technical documentation lists a 1.6B vision encoder. Use it to ask questions about supplied images or combine them with document text, with the answer returned as text.

Structured output and function calling are supported. Your application still supplies the tools, executes their calls and checks the result. The API is in public preview, so keep a working fallback when adding it to an existing workflow.

Create a Mistral API account, configure billing and put your key in the server environment as MISTRAL_API_KEY. This example sends a short text task directly to Mistral and caps generated output. Inspect the returned answer and usage before attaching a larger document or tool loop.

For the source in this example, check that the answer rejects the return and quotes “Opened software is excluded.” The 30-day window does not override that exclusion. Reject answers that invent exceptions or quote words absent from the source. Save the model ID, prompt, answer and returned usage together, including input and output token counts and any retries. Then send the same source and question to Small 4 with the same output cap and acceptance checks. Compare the cost of answers that pass those checks before expanding to longer documents; keep incorrect answers in the record so retries remain visible.

Mistral announced a 50% launch discount for two weeks from October 6, 2026. The rates below are USD per million tokens, checked on October 7. Use the standard column for ongoing budgets after the offer; the announcement does not specify the cutoff timezone.

For 1,000 requests averaging 20,000 uncached input tokens and 3,000 output tokens each, the token estimate is $19.87 at launch rates or $39.74 at standard rates. Formula: 20 million input tokens × input rate + 3 million output tokens × output rate.

Add the tokens used by retries and each tool-loop turn. Apply the cached rate only to eligible cached input reported by the provider; a repeated document is not automatically a cache hit. The figures exclude separately billed tools and provider fees.

For short classification or extraction jobs that fit its 256k context, Mistral Small 4 is a lower-cost candidate: its documented $0.15 input / $0.60 output rates give $4.80 for the same token workload. Large 4’s longer direct-API context matters when the relevant source material will not fit that window.

Artificial Analysis lists Mistral Large 4 Preview’s reasoning configuration at 38 on Intelligence Index v4.3.2 and 116.1 output tokens/s through Mistral’s API, accessed October 8, 2026. Output speed measures generation after the first chunk arrives; it excludes the wait before generation. AA lists a 524K context for this preview entry, while Mistral’s direct API documentation specifies 1M. Use the limit of your chosen route when sizing a request.

Choose on accepted results: preserve the same documents, output schema and correctness checks, and record retries as well as the initial answer. A higher token price can still be worthwhile if it avoids enough failed work. For repository tasks, compare with the model already integrated into your coding client before changing the whole workflow.

On October 6, the AI news account thehype. shared a haunted-room scene-code comparison. The author reported $0.61 and 72 minutes 36 seconds for Large 4, versus $0.04 and 30 minutes 9 seconds for DeepSeek V4 Pro, and described Large 4’s objects as the most detailed.

The post omits the endpoint, token usage and discount details. For your task, record those alongside retries and the time to an accepted result.

No. Active parameters describe the part used for a token; the full expert weights still need to be available to the inference system. Mistral’s current launch announcement reports 52B active parameters, and its technical documentation specifies 1.05T total. The active count does not reduce the amount of expert weights the serving system needs to store.

Using 1.05 trillion parameters × bits per weight ÷ 8 gives a raw-weight estimate of 2.1 TB at 16-bit, 1.05 TB at 8-bit or 525 GB at 4-bit, in decimal units. These numbers exclude quantization metadata, KV cache, runtime buffers and any separately stored components.

The launch announcement schedules weights for the end of October; the Hugging Face page is an upcoming-release listing. Use the hosted API now. A self-hosting purchase needs the released files, their license and a tested serving configuration before a hardware quote is meaningful.

For answers grounded in a document collection, use EmbeddingGemma 2 to index and retrieve relevant chunks, then supply those chunks to this model with their source identifiers. Keep retrieval quality and answer quality as separate checks: the correct passage can be found while the generated answer still misreads it.

Budget the first corpus embedding pass, later index updates, vector storage and generation inference separately. Changing embedding checkpoints requires a compatible new collection; lowering vector dimensions should be evaluated against labeled queries before it becomes a storage saving.