Mistral Large 4 (Le Chonk) on magicdoor.ai: Gateway Limits and Pricing

Direct answer: Mistral Large 4 (Le Chonk) is a public preview, not GA. The Mistral model reference lists a 1M-token model context, while the Vercel AI Gateway endpoint selected for the magicdoor.ai route caps context at 524,288 tokens and completion at 262,144 tokens. Use the Gateway cap for this route, not the direct API headline. Check the model directory and picker for current availability before starting a task.

Fact-checked October 8, 2026. Preview availability, endpoint limits, and sale pricing can change; check the current picker and linked sources before relying on this guide.

What the preview supports

The Gateway catalog lists text and image input, text output, tool calls, and structured output for mistral/mistral-large-4. The Magicdoor integration supports image input, search, Canvas, and reasoning but does not expose an explicit reasoning-effort control, PDF input, or code interpreter. Its conservative input budget is 262,144 tokens, lower than the 524,288-token Gateway context ceiling; the actual request also needs room for output. Use the model detail page to check the launched controls. Neither audio nor video input/output is listed for this endpoint.

Mistral's launch announcement calls this a public preview and says weights will be released later. Do not read “open-weight model” as evidence that downloadable weights have already shipped, or as a self-hosting feature in magicdoor.ai.

Gateway sale pricing, not a permanent rate

As checked October 8, 2026, both Mistral's model rate card and Vercel's billed endpoint metadata show these current sale rates in USD per million tokens:

These are usage rates, separate from magicdoor.ai's subscription. The sale can change; Magicdoor's guarded route checks the public Gateway endpoint before dispatch and fails closed if these guarded sale rates change until the rate card and billing are updated. Check the in-product quote before budgeting. Cached-input rates apply only to tokens that actually qualify. The general model-cost guide covers how input, output, and cached tokens differ; avoid assuming all prompt tokens are cached.

Privacy and route limits

Vercel currently lists one endpoint, upstream provider Mistral, with has_zdr=true and has_no_training=true. Eligibility metadata alone is not a ZDR guarantee; the Magicdoor route pins Mistral and sends zeroDataRetention: true on each request. It checks that the sole endpoint, retention flags, limits, and sale rates still match before dispatch; if they change or cannot be checked, the request fails closed. Vercel's ZDR documentation explains per-request enforcement. This is provider inference policy; normal magicdoor.ai conversation storage remains separate.

There is no listed GreenPT endpoint for Large 4. A general preference for renewable inference does not make this model run through GreenPT. For currently supported renewable routes, see renewable-powered AI.

Choosing and checking the model

Try the same text or image-grounded task with a second supported model and judge the responses against your own rubric; a provider benchmark is not a guarantee for your workflow. If your task needs more than the Gateway's 524,288-token context, split the source material or choose a different supported route. If privacy enforcement is a hard requirement, confirm the launched request path still enforces per-request ZDR before sending sensitive material; this does not change chat-storage behavior.

Browse all available models and the model-selection guide, then open chat to check whether the preview appears in your picker.

FAQ

Is Mistral Large 4 generally available on magicdoor.ai?

No. Mistral describes Large 4 (Le Chonk) as a public preview, not a generally available release. Availability through magicdoor.ai depends on its supported Vercel AI Gateway route; check the model picker before relying on it.

Is the context window one million tokens on magicdoor.ai?

No. Mistral documents a 1M-token direct model context, but the Vercel AI Gateway endpoint used by magicdoor.ai lists 524,288 tokens of context and up to 262,144 output tokens. The Gateway limit is the relevant ceiling here.

What are the current Mistral Large 4 token rates?

As checked on October 8, 2026, the Vercel Gateway endpoint lists sale rates of $0.68 per million input tokens, $0.07 per million cached input tokens, and $2.09 per million output tokens. These are time-sensitive rates, not a permanent price promise; check current model pricing before sending.

What privacy route does Magicdoor use for Mistral Large 4?

The Magicdoor route pins the Mistral endpoint through Vercel AI Gateway and requires per-request ZDR. Its pre-dispatch check fails closed if the endpoint, guarded sale rates, or retention eligibility change or become unavailable. This is provider inference policy, not a promise about magicdoor.ai chat storage. No GreenPT route is listed.