EuroInference compares the same AI model across European providers. Prices are public listings, Quality is deggo Quality v4 (via models.deggo.fyi), and EUR Value is EuroInference's heuristic (Quality × affordabilityEU) mixing that Quality with EU pricing.
What you see
* Input / Output — cheapest list price per 1M tokens. * Workload Cost — what you'd actually pay for your token counts, with your caching settings. Cheapest offer wins. * Quality — deggo Quality v4 (0–100) via models.deggo.fyi. * EUR Value — quality per euro at list price: Quality × affordability, where blended = (input + output)/2 *from the single cheapest EU offer* (the offer with minimal (in+out)/2; input and output are taken together, not mixed across providers) and affordability = 1/(1+log10(1+blended·8)·0.45) — 8 and 0.45 are heuristic scaling (log-dampened per Weber-Fechner) chosen to spread Value across our price range, not fitted to demand data. Higher = more quality for the money. Not affected by the workload sliders.
Switching EUR/USD only converts the displayed numbers — rankings stay on EUR prices.
Workload Cost in one line
p_eff = base·(1−s) + (s/R)·[w + (R−1)·r]
total = (p_eff·input + out·output) / 1,000,000* base / out — list prices. * s — share of input from cache (Chat 0%, RAG 40%, Agentic 80%). * R — how often a cached prompt is reused (2, 4, 8, or ∞). * r / w — provider's cache read / write price. If a rate isn't published, we use base (no discount) — unknown never means free.
Turn off *prompt caching* and p_eff = base.
Caching in practice
Defaults are Agentic 80% / R=8, on. That's intentionally below top agents (Claude Code 92% hits) and matches published measurements of 41–80% savings for agentic workloads.
Per offer we show: Priced (rate published), Supported, no rate, Not supported, Unknown. Only *Priced* gives a discount.
Where the numbers come from
Prices are fetched directly from the providers' own APIs twice daily from their public listings (Cortecs, EURouter, Mammouth, Mistral, etc.). Mistral @regional is 1.1×. Invalid rates are dropped.
Quality is deggo's published Quality v4. We don't store the raw benchmark tables.
EUR Value is our heuristic calculation: we take the cheapest EU offer *as a whole* — the offer with minimal (in+out)/2 in EUR, input and output taken together from that same offer — and apply the affordability curve above (heuristic, not an econometric model).
The OpenRouter card in the detail view is a non-EU reference for comparison only — never used for rankings or costs.
Sovereignty badges
Each provider card shows six rows — Jurisdiction · Retention · Training · Hosting · Inference · Routing — with a short badge plus a hover tooltip that quotes the source sentence from config/sovereignty.json (or a per-listing API signal). ? Unknown means the provider doesn't publish that fact — we never fabricate it.
Jurisdiction — legal home of the gateway operator. 🇫🇷🇪🇺 France / 🇳🇱🇪🇺 Netherlands = EU member (GDPR, EU enforcement), 🇬🇧 United Kingdom / 🇺🇸 United States = non-EU. Flag + 🇪🇺 when inEu: true.
Retention — what happens to prompts/completions after the call. ✓ Zero-day = provider states payload is processed in-memory only and not retained (e.g. Cortecs, EURouter, GreenPT API). ✗ Retained ≤ 30d = kept for abuse-monitoring/billing (Mistral 30d unless ZDR is activated, Mammouth 30d). Per-model overrides come from the catalog when present (e.g. Requesty data_retention_days).
Training — whether your prompts are used to train models. ✓ No = provider states no training on customer data (Cortecs, EURouter, GreenPT, Mistral, etc.). ✗ Possible = depends on the routed upstream model (OpenRouter true — some endpoints train). Again, Requesty exposes data_used_for_training per model when present.
Hosting — where the *gateway/control plane* runs (not the model). ✓ EU = EU data centre (e.g. Scaleway FR-PAR / NL-AMS, OVH FR, GCP Belgium). ✓ EU (US) = EU-located but operated by a US hyperscaler (AWS Stockholm, Azure, GCP) — CLOUD Act may apply despite EU location. Global = worldwide (Requesty, OpenRouter). ✓ EU-Sovereign = EU jurisdiction + EU hosting on non-US operator (e.g. Scaleway, OVHcloud). Tooltip cites the exact sentence and source.
Inference — where the *model itself* runs. ✓ EU = EU-only (Cortecs, EURouter, GreenPT on Scaleway FR). Global = no commitment — prompt may leave the EU (Mammouth lists Fireworks US / xAI US, Eden AI api.edenai.run global, Opper global unless EU route selected, Mistral global api.mistral.ai — no commitment, subprocessors include Google Cloud/Azure EU+US). ✓ EU (US) = EU region of a US provider (EU Bedrock/Vertex/Azure) — still under US law. ✓ EU-Sovereign = the strict badge: jurisdiction EU + hosting EU non-US + inference EU non-US + (European model OR open-weights hosted on non-US EU infra). Example: Mistral models on EU fleet qualify; the same model via a US host does not. The model-creator flag (Mistral AI = European, Meta/Alibaba/DeepSeek = open-weights) and usInferenceRe host signals are documented in app.js.
*Concrete example:* Mistral AI vs Mistral AI Regional. The global endpoint shows Inference: Global (tooltip: *“Global endpoint api.mistral.ai makes no region commitment; inference may run outside the EU (sub-processors include Google Cloud and Azure). Only api.eu.mistral.ai pins inference to the EU.”*). Mistral AI Regional shows Inference: ✓ EU-Sovereign (tooltip: *“Inference pinned to EU via api.eu.mistral.ai (1.1× surcharge); guarantees EU data residency.”*). Same company, same 30-day retention and no-training, but only the api.eu.mistral.ai offer is EU-Sovereign. Price reflects it (1.1×).
Routing — architecture. Direct = first-party fleet (Mistral). Aggregator = gateway in front of many upstream model providers (Mammouth, Cortecs, Eden AI, EURouter, GreenPT, Requesty, OpenRouter).
All sovereignty facts are resolved once at generation time (unify.js: resolveOfferSovereignty() + SOVEREIGNTY_META); the browser only merges the per-offer scalars with the shared SOVEREIGNTY_META sentences. Missing blocks stay ? Unknown and badges never fall back to EU by default. Sources are linked in config/sovereignty.json; verify the provider's DPA/terms for binding commitments.
Limits
Prices exclude VAT and are converted with the periodic ECB reference rate shown in the header (timestamp, typically twice daily — may lag live catalogs). Caching estimates depend on your prompts, assumptions (share/reuses), and the provider's TTL — only published positive rates reduce the estimate, unknown never means free, and formulas are documented above. Quality (deggo Quality v4) is a synthesized third-party benchmark and EUR Value is EuroInference's heuristic (Quality × affordabilityEU, affordability=1/(1+log10(1+blended·8)·0.45) over the cheapest EU blended (input+output)/2); neither is an official rating. Sovereignty badges summarize self-published provider policies and API signals (sources in config/sovereignty.json); omitted = unknown, not a certification — verify the provider's DPA and binding terms. The OpenRouter card is a non-EU reference for price comparison only, excluded from all rankings and calculations. Generic model facts are enriched when matched from models.dev (MIT). Always check the provider's own pricing page, limits, privacy, and terms before buying. Data is provided on an 'as is' basis without warranties beyond mandatory statutory limits.