이미지: METAL LAB 생성
Summary
- Mistral has officially launched regional endpoints letting customers designate EU or US servers, charging a 10% premium over standard pricing
- Regional endpoints support only function calls; agents, batch processing, and file management fall outside the regional guarantee
- A paid Priority Tier with a 99.5% uptime SLA also launched in public beta alongside the regional endpoints
- 지역 엔드포인트 주소
- api.eu.mistral.ai / api.us.mistral.ai (정식 출시)
- 지역 지정 추가요금
- 표준 요금 대비 10% 할증
- 지역 엔드포인트 지원 범위
- 함수 호출만 지원, 에이전트·배치처리·파일 관리 미지원
- 지역 제한 예외 항목
- 계정 설정·API 키·결제·사용 통계는 지역 밖 처리 가능
- 우선순위 등급(Priority Tier)
- 공개 베타, 가동률 99.5% SLA(월 약 3.5시간 다운타임 허용)
- 배경 발표
- 8월 11일 미스트랄 공식 X, European Compute Units·GLM-5.2 통합 계획 공개
Routing to European servers costs more
When a bank or government agency sent a contract to Mistral, there was previously no way to confirm which country's servers processed that data. Mistral has now started selling an answer to that question. But it comes at a price.
In a blog post on the 12th (local time), Mistral announced the official launch of regional endpoints that let customers direct requests to servers specifically in Europe or the US. Choosing between the EU endpoint (api.eu.mistral.ai) and the US endpoint (api.us.mistral.ai) locks processing to that region. Requests sent to the default endpoint carry no guarantee about where they're processed. Regional designation adds a 10% surcharge on top of standard pricing.
Between the word "sovereignty" and the reality
This regional processing doesn't extend to all the add-on features Mistral sells. The only capability that works through regional endpoints is function calling — the model's ability to call external APIs. Agents, batch processing, and file management are unavailable at the regional addresses. Model availability also varies by region, and rather than publishing a fixed list, Mistral requires customers to query each endpoint directly to check what's supported.
| Feature | Regional Endpoint | Default Endpoint |
|---|---|---|
| Function calling | Supported | Supported |
| Agents | Not supported | Supported |
| Batch processing | Not supported | Supported |
| File management | Not supported | Supported |
| Processing location guarantee | Guaranteed (10% surcharge) | Not guaranteed |
According to the documentation, account settings, API keys, billing, and usage statistics may be processed outside the selected region. The blog also mentions the possibility of limited, secured data transfers to subcontractors outside the region. In other words, only the compute step is localized — not the entire platform. Whether requests are subsequently stored or logged depends on a separate setting called "Zero Data Retention."
Why agents were left out
Simple model calls don't need to retain state. Agents, batch jobs, and file storage, by contrast, need to preserve things like intermediate steps or uploaded documents even after a call finishes. Mistral refers to these as "stateful" features. While the company hasn't confirmed this directly, the likely explanation is that such features require separate local infrastructure.
Paying to jump the queue
The second announcement is the Priority Tier, currently in public beta. Because all customers share the same data centers, response times slow down when requests pile up. The Priority Tier acts as a kind of express lane, processing paying customers' requests first during congested periods. Mistral said it's targeting use cases where latency translates directly into losses — customer service chatbots or factory production systems, for example. The tier comes with an SLA guaranteeing 99.5% uptime, which contractually allows for roughly 3 hours and 30 minutes of downtime per month.
Background: the next move in Mistral's European compute coalition
This announcement extends the European AI sovereignty strategy Mistral outlined via its official X account on August 11. At the time, Mistral unveiled a multi-year contract structure called "European Compute Units" designed to pool long-term compute demand within Europe, previewing both regional endpoints and a priority tier, and said it would bring third-party open models onto its platform starting with Zhipu AI's GLM-5.2. This report essentially maps out how far those two previewed features actually reach — and where they fall short.
How to use it
Developers using the Mistral API can apply regional routing simply by changing their request address to api.eu.mistral.ai or api.us.mistral.ai. 1) Point API calls to the desired regional endpoint address. 2) Query that endpoint directly to confirm which models it supports. 3) If your workflow needs agents, batch processing, or file management, design around the fact that you'll need to use the default endpoint for those. 4) If latency is a concern, sign up for the Priority Tier beta to enable that separate setting.
This applies to all enterprise customers using the Mistral API; regional endpoints carry a 10% surcharge, while the Priority Tier reportedly comes with separate contract terms. For example, a bank summarizing contracts by feeding them directly into the model without function calls could process that data via the EU endpoint without it leaving the region. But workflows that upload documents for file API management or that run agents remain outside the regional guarantee.
What actually changes
Companies adopting Mistral to satisfy European regulatory requirements now need to scrutinize exactly which features fall under "regional processing" rather than taking the phrase at face value. Customers who mainly use function calling get a meaningful new option, but those relying on file management or agents still have no way to guarantee where their data is located. The Priority Tier, meanwhile, is essentially a product for buying speed — a sign that competition in AI infrastructure is expanding beyond model performance to questions of where processing happens and how fast.
