OVHcloud AI Endpoints
ovh.net · Inference · 1 record · JSON
OVHcloud AI Endpoints (anonymous tier) Inference
★ qualifiesanonymousno cardMrequest-rate
| Free tier | no key required; 22 model ids on /v1/models, each carrying a pricing object in USD per token. Twelve of the 22 are chat models and two of those, Qwen3Guard-Gen-0.6B and Qwen3Guard-Gen-8B, price at exactly zero on prompt and completion. The other 10 ids are one image model, three embedding models, four TTS models and two ASR models unit tokens; metered on request-count; the pricing object is published per model and is the machine form of 'free'. Qwen3Guard-Gen-0.6B and Qwen3Guard-Gen-8B are the only two chat models priced at exactly zero on both prompt and completion. No chat model is priced at 0.00000001 USD: that figure belongs to the two embedding models bge-m3 and bge-multilingual-gemma2 |
|---|---|
| Price | none stated |
| Per-object cap | not applicable: an inference endpoint stores nothing for the caller |
| Retention | not applicable |
| Egress | the response is the payload; no separate egress term is published |
| Identity | anonymous no Authorization header is sent and none is required. The prober's per-endpoint track record reports 244 of 343 calls answered over 14 days from its own IP, 71 per cent, and 244 of the 267 calls OVH accepted for a model it serves, 91 per cent |
| Limits |
|
| Binding | request-rate |
| Interface | rest |
| Tooling | any OpenAI-compatible client; curl against POST {base}/chat/completions |
| Composes with | pairs with a free-standing agent as a complete no-account stack; the two-a-minute ceiling is the scarce link, not the agent |
| Reliability | the per-minute ceiling is 2, published in the response headers, so a burst is refused with 429 rather than throttled |
| Notes | The endpoint answers without a key but publishes a price per token, so 'keyless' describes the authentication requirement and not the cost. The two guard models are genuinely zero. Qwen3.5-397B-A17B and Qwen3.6-27B return empty content at max_tokens 256 with the reasoning field populated and finish_reason length, which is budget exhaustion and not a failure. |
| Evidence rung | measured eight models were called twice. The header ceiling was read from a live response. What is not measured is whether the zero price survives signup. |
| Checked | 2026-10-23 (live); stale on 2026-11-22 |
Evidence (2)
| date | source | what it proved |
|---|---|---|
| 2026-10-22 | probe:https://oai.endpoints.kepler.ai.cloud.ovh.net/v1/chat/completions | gpt-oss-20b answered 200 with the requested model echoed and the needle present, 51 completion tokens, 1305 ms (run 1) and 1241 ms (run 2) |
| 2026-10-22 | oai.endpoints.kepler.ai.cloud.ovh.net/v1/models | 22 model ids with context_length and a per-token pricing object, 12 of them chat models; Qwen3Guard-Gen-0.6B and -8B carry prompt 0 and completion 0; bge-m3 and bge-multilingual-gemma2 carry prompt 0.00000001 and are embeddings; the largest chat model, Qwen3.5-397B-A17B, carries prompt 0.00000071 and completion 0.00000425 USD per token at 262144 context |
Unknown or withheld values (10)
$.probe.records[0].tps_basisprocess$.probe.records[1].tps_basisprocess$.probe.records[2].tps_basisprocess$.probe.records[3].tps_basisprocess$.probe.records[4].tps_basisprocess$.probe.records[5].tps_basisprocess$.probe.records[6].tps_basisprocess$.probe.records[7].tps_basisprocess$.probe.records[8].tps_basisprocess$.retention.statednot-applicable
Corrections
- 2026-10-22 The prober's answer rates were right and the reading of them was wrong