Mistral, the Paris-based open-weight lab, has made its EU-only inference endpoint generally available, letting customers route model requests through servers that stay inside Europe. The guarantee stops well short of the platform’s full feature set: agents, batch processing, and file storage sit outside the regional promise. For any European enterprise treating this as a sovereignty checkbox, the exceptions determine whether the checkbox actually applies.
The service runs through two addresses, api.eu.mistral.ai and api.us.mistral.ai, each pinning compute to a single geography. Customers who stay on the default endpoint get no commitment about where a request is processed. Mistral charges a 10 percent premium over standard pricing for regional routing, and the shorter network path also lowers latency.
Regional endpoints support exactly one add-on capability: function calling, which lets a model invoke outside APIs. Agents, batch processing, and the Files API are unavailable there, and model selection differs by region because Mistral does not publish a fixed list. Because Mistral publishes no fixed roster, teams must poll each address directly to find out which models are currently live.
The likely explanation is architectural. A single inference call requires nothing be retained, but agents, batch processing, and file uploads all carry data forward through several steps, a property Mistral’s documentation labels “stateful.” Those probably require additional on-site infrastructure in each region, though the company has not confirmed that reasoning, as The Decoder noted in its reading of the rollout.
The regional guarantee, in practice, covers far less than the word “sovereignty” would suggest. Account details, API credentials, billing records, and usage metrics may still be handled beyond the selected region’s borders, and the documentation notes that Mistral allows narrowly scoped, secured handoffs to subcontractors elsewhere. Only the compute step itself is regional. A distinct toggle, called Zero Data Retention, determines whether the request gets saved or logged once processing finishes.
Mistral is also opening a Priority Tier in beta, a paid fast lane that moves a customer’s requests ahead of standard traffic during congestion. The company is pitching it at latency-sensitive uses like customer service chatbots or factory floor production systems. The tier’s service level agreement guarantees 99.5 percent uptime, translating to a downtime allowance of about three and a half hours per month, a floor the standard tier does not offer.
Customers turn on priority routing with an API parameter called service_tier, set to “auto” for the fast lane or left at the default “standard_only.” Rate limits for priority traffic are negotiated per customer, and requests over that limit fall back to standard processing rather than failing. Priority access costs 1.75 times the standard rate, with prompt caching discounts, which can reach 90 percent, calculated first. The tier is not self-service and requires a contract with Mistral’s sales team.
Separately, Mistral is opening its platform to third-party open models, starting with GLM-5.2 from Z.ai, a Chinese AI company, which will run under the same regional rules as Mistral’s own models. To fund the added compute, Mistral is asking large customers for multi-year purchase commitments it calls European Compute Units, a bet that new European data center capacity only gets built if enough customers sign on long term.
Before signing anything, European buyers should ask Mistral one specific question: which request paths, agents, file uploads, batch jobs, account management, billing, all of it, actually leave the EU boundary, and under what conditions the subcontractor exception kicks in. A residency guarantee that covers only the model call is a procurement detail, not the sovereignty story Mistral is selling.
The Decoder, the German AI news outlet, reported these platform changes on August 12, 2026, based on Mistral’s own documentation for regional inference and the Priority Tier.
