Mistral would rather sell you the pipes than the model

Mistral said Tuesday that it will begin hosting third-party open models, starting with GLM-5.2 from the Chinese lab Z.ai ^[2]. The model runs on the same infrastructure as Mistral's own, with access to the company's regional processing controls and its new priority service tier ^[3], which tells you where Mistral thinks the durable business is: not in winning every model bake-off, but in being the place the bake-off happens.

The commercial logic is straightforward for a company whose reputation was built on releasing open-weight models ^[1]. If enterprises are going to swap models every few months, the vendor that owns the endpoint keeps the account. Mistral's pitch is one place to run different open models without starting over on each switch ^[4].

The specifics. GLM-5.2 has a 1 million-token context window, and Mistral lists coding and long-context agentic work among its main uses ^[5]. It is available as zai-glm-5-2 at $1.40 per million input tokens, $4.40 per million output tokens, and $0.14 per million cached input tokens ^[6]. Regional endpoints, which let developers pin processing to Europe or the United States by changing the base URL, are now largely available in both ^[7][8]. That guarantee costs 10% on every input, output and cached token ^[9], so a European-pinned GLM-5.2 call works out to roughly $1.54 in and $4.84 out per million ^[10].

Now the conditions. The regional endpoints do not support everything: function calling works, Agents, Batch and the Files API do not, and some models are missing depending on region ^[11]. If your application depends on an unsupported feature, the regional processing guarantee stops covering that workload ^[12], which means migration is more than a URL change. Mistral also says some account and usage data may still leave the selected region, and information may be shared with outside companies under the safeguards in its Trust Center ^[13]. The Microsoft-Mistral sovereign compute partnership answers some of this for Azure customers; teams on Mistral's own endpoints get a different set of guarantees ^[14].

The Priority Tier, in public preview, puts eligible calls ahead of Standard Tier traffic and carries a 99.5% uptime SLA ^[15]. Eligibility is narrower than the label suggests. Access must be arranged with Mistral, with custom limits agreed per model, and the request needs `service_tier: auto` set explicitly ^[16]. A call is only prioritised if the entitlement is active, the model is covered, the request is inside its custom rate limit, and Mistral has capacity for that model in that region; miss one and it silently drops to Standard ^[17]. Mistral does report which tier served each request in the response's usage object ^[18], and that field is the only honest instrumentation available here. Log it.

One more thing worth saying plainly: open weights hosted by Mistral is still a hosted service. Mistral decides which version is available and runs the infrastructure behind it ^[19]. And a shared API does not make models interchangeable, so each one still needs its own evaluation before production ^[20].

What to watch: whether the downgrade rate on the priority tier stays low enough to justify the entitlement negotiation, whether Agents and Batch reach the regional endpoints, and which model Mistral adds second. The second addition tells you whether this is a strategy or a one-off.