Nebius Token Factory review: an inference provider with a clear path to dedicated capacity
Answer in brief: Nebius Token Factory ranks #8 among the best AI inference providers. It is our pick for teams that prototype on shared endpoints, then move to isolated dedicated endpoints in EU or US regions. Serverless comes in Fast and Base flavors behind an OpenAI-compatible API. Dedicated endpoints carry a 99.9% SLA and are billed per GPU-hour at per-minute granularity, with no charge at zero replicas. Watch the serverless catalog: 11 models leave Serverless on November 9, 2026.
Independent research, checked 2026-11-09 against Nebius' product page, GLM model page, documentation and deprecation notice. We did not test the product hands-on.
Start shared, graduate to dedicated
Token Factory is Nebius' inference and post-training platform. You start with public serverless endpoints, choosing Fast for low-latency interactive work or Base for cost-efficient bulk processing, on the same API. When you need stable versions, isolation or guaranteed capacity, you move to dedicated endpoints. You can also upload and host LoRA or fully fine-tuned models through the dashboard or API, use built-in embeddings for RAG, and run code in Sandboxes.
Who Nebius Token Factory fits best
Nebius fits teams with EU or US data-residency requirements, companies scaling one model to very high volume on dedicated capacity, and anyone who wants a single vendor from playground experiments to isolated production endpoints.
Why teams pick Nebius Token Factory
- Dedicated endpoints done right: a 99.9% SLA, custom autoscaling and optional regional deployment in EU or US data centers.
- Fair dedicated billing: per GPU-hour with per-minute granularity. Billing only applies while at least one replica runs, and provisioning, restarts and graceful shutdown are not billed.
- Competitive serverless prices: GLM-5.3 $1.40 / $4.40 and GLM-5.3-Flash $0.15 / $0.50 per 1M tokens, with automatic cached-token discounts on Serverless.
- Security: a zero-retention mode, plus SOC 2 Type II, HIPAA and ISO 27001. Data centers are in Finland, France and the US.
- Production proof on Nebius' own site: Nebius quotes Prosus as handling up to 200 billion tokens per day on dedicated endpoints with autoscaling.
- Investment in the stack: Nebius acquired Inferize on October 1, 2026; its team and technology joined Token Factory.
The honest drawbacks
- Serverless models get retired. Nebius' November 2026 notice removes 11 models from Serverless on November 9, 2026: GLM-5.1, NVIDIA-Nemotron-3-Nano-30B-A3B, nemotron-3-super-120b-a12b, DeepSeek-V4-Pro, DeepSeek-V4-Flash-0731, DeepSeek-V4.1-Flash, Qwen3.5-397B-A17B, Qwen3-235B-A22B-Instruct-2507, Kimi-K2.6, Kimi-K2.7-Code and MiniCPM-V-4_5. Requests are not rerouted automatically.
- Nebius points production workloads that need stable model versions to dedicated endpoints.
- The full self-service price catalog requires a console login.
- Cached-token discounts apply only to Serverless; dedicated endpoints stay on GPU-hour billing.
Pricing at a glance
| Item | Published price |
|---|---|
| GLM-5.3 (Serverless) | $1.40 input / $4.40 output per 1M |
| GLM-5.3-Flash (Serverless) | $0.15 / $0.50 per 1M |
| gpt-oss-120b (Serverless) | $0.15 / $0.60 per 1M (price catalog, login required) |
| Dedicated endpoints | Per GPU-hour, per-minute granularity; varies by GPU and region |
Best Nebius Token Factory alternatives
- Together AI: a broader serverless catalog plus Provisioned Throughput.
- Baseten: dedicated deployments for custom models and region compliance policies.
- DeepInfra: low serverless list prices across 150+ models.
Nebius Token Factory FAQ
Which models left Nebius Serverless in November 2026?
Eleven models were scheduled for removal on November 9, 2026, including GLM-5.1, DeepSeek-V4-Pro and Kimi-K2.6. Nebius lists a recommended replacement for each, and dedicated endpoints are unaffected.
How are Nebius dedicated endpoints billed?
Per GPU-hour, with per-minute granularity, while at least one replica is running. Zero replicas are not billed.
Is Nebius Token Factory OpenAI-compatible?
Yes. You point the OpenAI client at api.tokenfactory.nebius.com/v1 with a Nebius API key.
Sources
- Nebius Token Factory inference service
- Nebius GLM models inference page
- Model Deprecations: November 2026
- Dedicated endpoints overview
- Dedicated endpoint billing policy
- Cached tokens
- Token Factory quickstart
- Token Factory price catalog (login required)
- Nebius acquires Inferize (October 1, 2026)
← Back to the ranked list of the best AI inference providers
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.