Tested, not claimed
Marketing pages assert; this page proves. Every number below is either computed live from the model registry at build time, or tied to a dated measurement described in the code. When a number ages, the page ages with it — nothing here is hardcoded marketing copy.
Live counters
What the catalog actually contains
These counters are computed from the model registry when the page is built — change the registry, the numbers change. No manual sync, no stale figure.
27
live chat models
14
on EU sovereign infrastructure
15/27
carry a Frontière Verified checklist
10
benchmark-tracked model families
Verification detail
Every verified model, its provider, its score
Models carrying at least one verified criterion. A 6/6 score means all checks passed; a partial score shows exactly which criterion is still open.
| Model | Effective provider | Checklist |
|---|---|---|
| qwen3.5-397b | scaleway | 6/6 |
| glm-5.2 | scaleway | 6/6 |
| deepseek-v4-flash-0731 | scaleway | 6/6 |
| llama-3.3-70b | ovhcloud | 5/6 |
| qwen3.6-27b | ovhcloud | 6/6 |
| qwen3.5-9b | ovhcloud | 6/6 |
| qwen3-32b | ovhcloud | 6/6 |
| qwen3-coder-30b | ovhcloud | 6/6 |
| qwen2.5-vl-72b | ovhcloud | 5/6 |
| gpt-oss-120b | ovhcloud | 6/6 |
| gpt-oss-20b | ovhcloud | 6/6 |
| mistral-small-3.2-24b | ovhcloud | 6/6 |
| kimi-k3 | modal | 5/6 |
| qwen3.8-27b | ovhcloud | 5/6 |
| deepseek-v4.1-flash | modal | 6/6 |
Dated evidence
Each claim, its evidence, its date
Function calling works on the catalog
2026-08-06Measured, never inferred: one real API call per entry with a get_weather tool, plus a full role:"tool" round-trip. Result: 12 of the 13 models live at test time returned a valid tool_call with well-formed JSON arguments. The single refusal (qwen2.5-vl-72b, OVHcloud answers HTTP 400 "tool calls not supported") is documented on its model page.
See toolCalling field documentation →Zero prompt / completion retention
2026-07-29By design, not policy-only: the gateway stores neither prompts nor completions from API calls — only token counts required for billing. The processing is simple enough to be described in a short DPA (Art. 28 GDPR).
How we determine sovereignty →Sovereign models run on EU-controlled providers
2026-07-29OVHcloud and Scaleway — French companies with no non-EU capital control — plus our own dedicated EU servers (FLEECE AI SASU, French). Provider facts are transcribed from audits consigned in the registry; the effective provider per model is the priority-1 routing entry.
Sovereignty methodology →Published prices match provider prices plus a stated margin
2026-08-13Provider prices re-confirmed by direct checks (Scaleway pricing page, OVHcloud /v1/models endpoint, OpenRouter /v1/models) and converted at the documented ECB-referenced rates. maxCompletionTokens limits measured by real oversized requests per provider.
Registry source comments →Benchmark scores are real, public figures
2026-08-26MMLU, GPQA Diamond, LiveCodeBench and Arena Elo scores are the best publicly reported figures per model, sourced from Hugging Face model cards, official technical reports and LMSYS — not our own run, and labeled as such.
Benchmark data file →The score
Sovereignty Score — deterministic, recomputable
The Sovereignty Score shown on every model page is a deterministic 0–100 aggregation of five weighted factors computed from registry facts. You can recompute any score yourself from the published formula — no black box.
Read the full methodology →The source
One registry, every fact
Everything on this site — prices, jurisdictions, verification checklists, tool-calling results, context windows — comes from a single typed model registry whose comments document each verification: what was tested, when, how. It is the opposite of a marketing layer: the code is the source of truth.