New Provider: EURouter

[dfigos-2000]

Website: https://www.eurouter.ai/
API Docs: https://www.eurouter.ai/docs/api/chat
Models: https://www.eurouter.ai/models
It is OpenAI-compatible?: According to the docs, yes.

[diegosouzapw]

Hey @dfigos-2000! Same as GreenPT — good format, docs links included, and not on docs/reference/REMOVED_PROVIDERS.md, so nothing blocks it.

EURouter being OpenAI-compatible makes it a registry entry in src/shared/constants/providers/apikey/gateways.ts with passthroughModels: true — the mnn-ai / meganova-ai shape. No executor or translator work.

One thing about EURouter specifically, which does not apply to GreenPT: it is a router, not an inference provider. That has a consequence worth naming before anyone treats it as a sovereignty win — a router in the EU is a routing layer in the EU, and the model behind any given request may execute anywhere, under whatever terms that upstream sets. So "EU router" and "EU inference" are different guarantees, and the second is the one your original thread was actually about.

That is not an argument against adding it. It is an argument for the entry carrying honest metadata rather than implying more than it delivers, and for #12938's training flag needing a per-upstream answer here rather than one value for the whole gateway. If EURouter publishes which providers it routes to and under what terms, that link is the useful thing to attach.

Same three verification items before it lands:

1. Exact /v1 base URL, and whether GET /v1/models works with a key (required by passthroughModels: true).
2. Streaming SSE conformance — the usual place "OpenAI-compatible" gateways diverge, and never documented when they do.
3. Free tier specifics for freeNote, only from published sources.

A curl of /v1/models plus one streaming completion (key redacted) answers 1 and 2 immediately if you have an account.

[dfigos-2000]

/v1/chat/completions: This requires an API key

curl https://api.eurouter.ai/api/v1/chat/completions \
        -H "Authorization: Bearer $EUROUTER_API_KEY" \
        -H "Content-Type: application/json" \
        -d '{
      "model": "mistral-large-3",
      "messages": [
        {"role": "user", "content": "Say hello in three languages!"}
      ]
    }'
{"error":{"code":402,"message":"Insufficient balance. Please add credits to your account.","type":"invalid_request_error","param":null,"details":{"current":"0","limit":"0.005822296326"}},"requestId":"6ceda167-1669-4555-988e-0fb504797a77"}⏎

/v1/models: This one is shorted out because the response is quite big. Does not require an API key.

curl -X GET "https://api.eurouter.ai/api/v1/models"
{"object":"list","data":[{"id":"qwen3.6-27b","object":"model","canonical_slug":"alibaba/qwen3.6-27b","hugging_face_id":"Qwen/Qwen3.6-27B","name":"Qwen3.6 27B","created":1782400279,"description":"Qwen3.6 27B is Alibaba's multimodal Qwen3.6 model for vision, chat, reasoning, and tool use.","context_length":262000,"architecture":{"modality":"text+image->text","input_modalities":["text","image"],"output_modalities":["text"],"tokenizer":"Qwen3","instruct_type":null},"top_provider":{"context_length":262000,"max_completion_tokens":null,"is_moderated":false},"per_request_limits":null,"supported_parameters":["temperature","top_p","max_tokens","stop","tools","tool_choice","response_format"],"supported_voices":null,"knowledge_cutoff":null,"release_date":"2026-04-22","last_updated":"2026-04-22","reasoning":{"mandatory":false,"supported_efforts":[],"supports_max_tokens":false},"expiration_date":null,"links":{"details":"/api/v1/models/alibaba/qwen3.6-27b/endpoints"},"supported_api_endpoints":["chat.completions"],"default_parameters":{"temperature":null,"top_p":null,"top_k":null,"frequency_penalty":null,"presence_penalty":null,"repetition_penalty":null},"providers":[{"name":"Regolo","slug":"regolo"}],"pricing":{"prompt":"0.0000005","completion":"0.0000021","request":"0","image":"0","image_token":"0","image_output":"0","audio":"0","input_audio_cache":"0","web_search":"0","internal_reasoning":"0","input_cache_read":"0","input_cache_write":"0","discount":1,"currency":"EUR"},"author":"alibaba","slug":"qwen3.6-27b","tags":["open-source","vision","multilingual","long-context"],"author_info":{"slug":"alibaba","name":"Alibaba","display_name":"Alibaba (Qwen)","icon_url":"/images/logos/qwen.webp","gradient_from":"from-purple-600","gradient_to":"to-indigo-700","gradient_via":null,"website_url":"https://qwenlm.github.io"}},{"id":"pixtral-12b","object":"model","canonical_slug":"mistral/pixtral-12b","hugging_face_id":"mistralai/Pixtral-12B-2409","name":"Pixtral 12B","created":1765024302,"description":"Pixtral 12B is Mistral's multimodal model for image understanding, visual question answering, and document analysis.","context_length":128000,"architecture":{"modality":"text+image->text","input_modalities":["text","image"],"output_modalities":["text"],"tokenizer":"Tekken","instruct_type":null},"top_provider":{"context_length":128000,"max_completion_tokens":4000,"is_moderated":false},"per_request_limits":null,"supported_parameters":["temperature","top_p","max_tokens","stop","seed","response_format","tools","tool_choice"],"supported_voices":null,"knowledge_cutoff":null,"release_date":"2024-09-01","last_updated":"2024-09-01","reasoning":{"mandatory":false,"supported_efforts":[],"supports_max_tokens":false},"expiration_date":null,"links":{"details":"/api/v1/models/mistral/pixtral-12b/endpoints"},"supported_api_endpoints":["chat.completions"],"default_parameters":{"temperature":null,"top_p":null,"top_k":null,"frequency_penalty":null,"presence_penalty":null,"repetition_penalty":null},"providers":[{"name":"Scaleway","slug":"scaleway"}],"pricing":{"prompt":"0.0000002","completion":"0.0000002","request":"0","image":"0","image_token":"0","image_output":"0","audio":"0","input_audio_cache":"0","web_search":"0","internal_reasoning":"0","input_cache_read":"0","input_cache_write":"0","discount":1,"currency":"EUR"},"author":"mistral","slug":"pixtral-12b","tags":["open-source","vision","multimodal","instruction-tuned"],"author_info":{"slug":"mistral","name":"Mistral","display_name":"Mistral AI","icon_url":"/images/logos/mistral.jpeg","gradient_from":"from-[#ff7000]","gradient_to":"to-[#ff9a00]","gradient_via":null,"website_url":"https://mistral.ai"}},{"id":"glm-5.2","object":"model","canonical_slug":"zhipu/glm-5.2","hugging_face_id":"zai-org/GLM-5.2","name":"GLM 5.2","created":1781703617,"description":"GLM-5.2 is Z AI's long-context model for reasoning, coding, tool use, and multilingual chat.","context_length":1000000,"architecture":{"modality":"text->text","input_modalities":["text"],"output_modalities":["text"],"tokenizer":"ChatGLM","instruct_type":null},"top_provider":{"context_length":1000000,"max_completion_tokens":131072,"is_moderated":false},"per_request_limits":null,"

Streaming SSE conformance

EUrouter supports OpenAI-style SSE streaming via stream: true.

Feature	Status	Notes
stream: true	✅ Supported	Enables SSE streaming
Server-Sent Events	✅ Supported	OpenAI-style streaming
data: [DONE]	✅ Documented	Stream termination marker
choices[].delta.content	✅ Supported	Documented in SDK streaming examples
stream_options	⚠️ Undocumented details	Exposed by the API schema, but exact streaming semantics are not documented
Usage in final SSE chunk	⚠️ Undocumented	No published example establishing include_usage / final usage-chunk behavior
Tool-call streaming deltas	⚠️ Undocumented	Tool calling is supported, but exact SSE delta semantics are not documented
Reasoning streaming	⚠️ Undocumented	Reasoning controls exist, but streamed event semantics are not documented
Streamed error format	⚠️ Undocumented	Disconnect/retry behavior is documented, but no formal streamed-error event schema is specified

Assessment: EUrouter is OpenAI-compatible at the basic SSE protocol level. Several edge-case SSE behaviors are undocumented, rather than demonstrably incompatible. In particular, the published documentation does not provide enough raw SSE examples to verify usage chunks, tool-call deltas, reasoning deltas, or streamed errors.

Free tier

According to the published pricing documentation:

• Platform fee: €0/month
• Monthly request limit: 10,000 requests/month
• Daily limit: Unlimited
• Rate limit: 60 requests/minute
• Markup: 15%
• Model usage: Not free — token/model usage is charged separately according to model pricing

Important: "Free" refers to the platform plan; it does not mean unlimited free model inference.

Documentation inconsistency

There is a discrepancy in EUrouter's published documentation:

• Pricing page: 60 requests/minute
• Authentication documentation: 20 requests/minute

The pricing page states that it is the source of truth for current plan limits, so 60 RPM should be recorded for the Free tier, with the 20 RPM figure flagged as stale/inconsistent.

Summary

Property	EUrouter
OpenAI-compatible API	✅ Yes
SSE streaming	✅ Yes
[DONE] termination	✅ Documented
Full SSE edge-case conformance	⚠️ Partially documented
Free platform plan	✅ €0/month
Free requests	10,000/month
Free rate limit	60 RPM*
Free model inference	❌ No
Markup	15%

* EUrouter's authentication docs currently mention 20 RPM, while the pricing page states 60 RPM and identifies the pricing page as the source of truth.

Sources

• https://www.eurouter.ai/
• https://www.eurouter.ai/docs/api/chat
• https://www.eurouter.ai/models
• https://www.eurouter.ai/pricing
• https://www.eurouter.ai/docs/guides/streaming
• https://www.eurouter.ai/docs/concepts/authentication
• https://www.eurouter.ai/docs/concepts/credits

[diegosouzapw]

@dfigos-2000 — real curl output beats documentation every time. Opened #12985, and I ran the endpoint myself to extend what you found.

Confirmed and expanded:

Check	Result
GET /api/v1/models	HTTP 200, 274 KB, 147 models, no auth needed
GET /v1/models	HTTP 200 — the short path works too, so the base URL can simply be https://api.eurouter.ai/v1
Your /v1/chat/completions 402	Auth was accepted; the account just had no credits. That is a working key, not a rejected one.

The catalog is OpenRouter-shaped — canonical_slug, hugging_face_id, architecture, top_provider, pricing, per_request_limits, supported_parameters, providers, tags — so passthroughModels: true is viable.

Now the part that matters for why you opened the original thread, and the catalog answers it directly rather than leaving it to speculation:

"id": "claude-sonnet-5",
"canonical_slug": "anthropic/claude-sonnet-5",
"providers": [{ "name": "AWS Bedrock", "slug": "aws-bedrock" }]

Claude on EURouter is served through AWS Bedrock. Across the 147 models the upstream owners include openai (19), anthropic (9) and amazon (7), alongside mistral (24), alibaba (21) and zhipu (18).

So the caveat I raised is not a hypothetical: an EU router is a routing layer in the EU, and inference for a given model can run wherever that upstream runs. #12985 requires the entry's apiHint to state that plainly, so nobody reads "EURouter" as EU inference.

There is a genuinely good outcome buried in that, though. The per-model providers[] array is exactly the granularity #12938's training flag needs — one value for the whole gateway would be wrong here, and this is the concrete case proving it. I have linked it there.

Remaining gap before it can merge: streaming SSE was never exercised, since the 402 stopped the request before a stream existed. That is the usual place "OpenAI-compatible" gateways diverge, and it fails quietly. If you ever put credits on that account, one streaming completion would close the last open question.