> OVHcloud
OVHcloud AI Endpoints is a serverless, OpenAI-compatible inference API for 40+ open-weight models, hosted in Gravelines, France, and billed per million tokens per model with input and output priced separately. Published catalogue rates include gpt-oss-20b at EUR 0.04 per 1M input and EUR 0.15 per 1M output, gpt-oss-120b at EUR 0.08/EUR 0.40, Meta-Llama-3.3-70B-Instruct at EUR 0.67/EUR 0.67 and Qwen3.5-397B-A17B at EUR 0.60/EUR 3.60, with embeddings at EUR 0.01 per 1M input tokens and whisper-large-v3 at EUR 0.00004083 per second of audio. Several models are listed as Free in the catalogue (Qwen3Guard-Gen-8B and Qwen3Guard-Gen-0.6B, stable-diffusion-xl-base-v10 and the nvr-tts voices), and a Batch tier in beta is documented at 50 percent savings for bulk processing.
// pros
- Rate limits are published and tiered: anonymous access allows 2 requests per minute per IP and per model, while an API access key raises this to 400 requests per minute per Public Cloud project and per model, returning HTTP 429 on breach.
- Input and output token prices are itemised per model, ranging from EUR 0.04 per 1M input (gpt-oss-20b) to EUR 0.91 per 1M input and output (Qwen2.5-VL-72B-Instruct), so cost per model is comparable at a glance.
- The Batch tier (beta) is documented at 50 percent savings for bulk or deferred work, with up to 200 MB per batch or 50k rows and a 24-hour response SLO.
- Privacy and retention are explicit on the product page: zero data retention, with only the data required for billing kept, and customer data never used to train or improve OVHcloud's models; the service is covered by the Public Cloud Special Conditions, Appendix 10.
- Model lifecycle is transparent, with email and Discord notification to users active in the last 3 months and a 3-month grace period before removal; the published table lists decommission dates including 2026-09-16, 2026-10-01 and 2026-11-04.
// cons
- There is no credit grant: free usage is limited to specific models listed at zero price (guardrail models, SDXL base v1.0 and the NVIDIA Riva TTS voices), so mainstream LLMs are billed from the first token.
- The Batch tier is in beta and explicitly carries no SLA while in beta, so bulk-cost savings come with reliability caveats.
- The 400 requests-per-minute ceiling is per project and per model, and the docs note that a token- or character-based usage limit feature is only being considered, so there is no built-in spend cap.
- Published model decommissions run on a quarterly cadence (for example Qwen3-32B on 2026-09-16 and Mistral-7B-Instruct-v0.3 on 2026-10-01), so pinned model IDs can start returning 404 within months.
OVHcloud AI Endpoints is the stronger option where European data residency, zero retention and per-model EUR token pricing matter, and its OpenAI-compatible API means it can sit behind the same client as any other per-token provider in this directory; use the Batch tier for large offline classification or labelling jobs and keep a fallback model ID for each workload to absorb the quarterly decommission schedule.
written by Mike Fleming, Uprouter editorial · verified against official sources September 16, 2026
- OVHcloud AI Endpoints model catalogue (per-million-token EUR rates, Free models, context sizes) ↗
- OVHcloud AI Endpoints product page (40+ models, zero data retention, Base/Batch tiers, 400 requests per minute) ↗
- OVHcloud docs - AI Endpoints billing and model lifecycle (decommission table, grace period) ↗
- OVHcloud docs - AI Endpoints features, capabilities and limitations (rate limits, payload limits) ↗
Free tier present — value not quantified
This provider offers a free tier, but its terms don’t convert cleanly into a USD figure (e.g. request-count limits or capacity-dependent pools). We refuse to print a misleading zero here. See the note below and the official pricing page for the exact limits.
note: No credit or allowance stated on official pages (checked 2026-09-16); the official AI Endpoints catalogue instead lists specific models as Free (Qwen3Guard-Gen-8B and -0.6B, stable-diffusion-xl-base-v10, and the nvr-tts-de-de/it-it/en-us/es-es voices) with all other models billed per million tokens.
| Plan | Type | Monthly | Input /1M | Output /1M | Note |
|---|---|---|---|---|---|
| Free tier (source-reported) | free | — | — | — | Has a free tier: call for free (rate-limited) with no API key (or an empty key); registered users get higher rate limits. Totals subject to change by policy. |
| Free tier | free | — | Free | Free | rate-limited on new models |
| Pay-as-you-go | payg | — | $0.200 | $0.600 | Baseline rate for Llama and DeepSeek models. |
Pricing normalized from public sources — always verify with the provider.
// model prices
$ Does OVHcloud have a free tier?
Yes — OVHcloud offers a free tier. Always confirm current limits on OVHcloud's official pricing page — free tiers change without notice.
$ How much does OVHcloud cost per 1M tokens?
OVHcloud does not publish simple per-1M-token pricing (it may bill per request, per output, or via a subscription plan). Check their official pricing page for exact figures.
$ Is OVHcloud safe to use?
Uprouter rates OVHcloud as low risk. Risk is low on data handling and moderate on model continuity. The product page states zero data retention with only billing data kept and no use of customer data for model training, hosting is in Gravelines, France, and the service sits under OVHcloud's Public Cloud Special Conditions Appendix 10 - all verifiable official commitments. The commercial terms are equally explicit: no credit grant, per-token billing from the first request, and a 400 requests-per-minute ceiling per project and model. The residual risk is lifecycle churn rather than policy, with published decommission dates roughly three months after announcement and no SLA on the beta Batch tier. Risk ratings are editorial, evidence-based, and never influenced by affiliate relationships.
Raw community sentiment, shown unblended. Votes never feed into objective data fields like price, risk, or status — and never into sort order.
No published reviews yet — be the first.
// more_informationprice history · live status · risk & confidence · code snippets
price_history
No recorded changes yet.
live_status
full history →best-effort probes · not a guarantee · last 15h ago
| Checked | Status | Latency |
|---|---|---|
| 15h ago | up | 153ms |
| 1d ago | up | 177ms |
| 3d ago | up | 239ms |
| 4d ago | up | 80ms |
| 6d ago | up | 114ms |
| 7d ago | up | 265ms |
| 9d ago | up | 176ms |
| 10d ago | up | 235ms |
| 12d ago | up | 258ms |
| 13d ago | up | 146ms |
| 14d ago | up | 158ms |
| 15d ago | up | 228ms |
| 15d ago | up | 110ms |
| 16d ago | up | 81ms |
| 17d ago | up | 251ms |
| 18d ago | up | 111ms |
| 19d ago | up | 185ms |
| 19d ago | up | 129ms |
| 20d ago | up | 129ms |
| 21d ago | up | 163ms |
| 22d ago | up | 100ms |
| 22d ago | up | 187ms |
| 23d ago | up | 152ms |
| 24d ago | up | 189ms |
| 25d ago | up | 103ms |
| 26d ago | up | 155ms |
| 26d ago | up | 260ms |
| 27d ago | up | 144ms |
| 28d ago | up | 180ms |
| 29d ago | up | 293ms |
Uprouter runs its own probes; best-effort, not a guarantee. Probe results depend on our network location and cadence — for production, always use the provider’s own status page.
risk_&_confidence
Risk is low on data handling and moderate on model continuity. The product page states zero data retention with only billing data kept and no use of customer data for model training, hosting is in Gravelines, France, and the service sits under OVHcloud's Public Cloud Special Conditions Appendix 10 - all verifiable official commitments. The commercial terms are equally explicit: no credit grant, per-token billing from the first request, and a 400 requests-per-minute ceiling per project and model. The residual risk is lifecycle churn rather than policy, with published decommission dates roughly three months after announcement and no SLA on the beta Batch tier.
Confidence reflects how much verifiable evidence (official docs, probe history, community reports) backs this entry. Risk is our editorial assessment of operator reliability and terms — not financial advice.
code_snippets
# OVHcloud's own API — get a key from their dashboard
curl https://YOUR-PROVIDER-ENDPOINT/chat/completions \
-H "Authorization: Bearer THEIR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen",
"messages": [{ "role": "user", "content": "Hello" }]
}'Direct to the provider — Uprouter receives nothing; verify their current pricing before purchasing.
Ranked by neutral similarity signals only — live status, free-tier shape and free-credit proximity. Commercial relationships never influence this list.