Nebius Token Factory Review: Pricing & Comparison
Nebius's open-model inference platform (formerly AI Studio) publishes a $-per-million-token price and a measured throughput figure per model
Last verified: 2026-10-15 · Visit official site →
An open-model inference platform renamed from AI Studio
Nebius Token Factory is the inference product of the European GPU cloud vendor Nebius, previously called Nebius AI Studio (the marketing page still lives at nebius.com/ai-studio, the console at tokenfactory.nebius.com, and the OpenAI-compatible endpoint is api.tokenfactory.nebius.com/v1/). Like Baseten it belongs to the “hosted open-weight models” category: no GPT, Claude or Gemini here.
What is unusual about it is that price and measured throughput are published together, rather than just a unit price.
Shown on the console homepage (USD per 1M tokens, input/output; throughput as stated by the vendor):
- GLM-5.3-Flash: $0.15 / $0.50, 349 tok/s (us-north1)
- DeepSeek V4.1 Flash: $0.30 / $1.20, 139 tok/s (us-north1)
- DeepSeek-V4-Pro-0813: $1.32 / $3.96, 37 tok/s (uk-south1)
Underlying GPU cloud (per GPU-hour): HGX B300 $9.50, HGX B200 $8.50, HGX H200 $5.40, HGX H100 $4.50Mind the page’s own caveat: prices are approximate and depend on actual usage, and the full price list is behind login — so a complete cost calculation is not possible from outside.
Coverage and engineering scope
The featured models are the DeepSeek V4 Pro, DeepSeek V4.1 Flash, Z.ai GLM-5.3-Flash, MiniMax M3 and Kimi K3 tier of open models; the full catalogue requires a login and no total count is published.
As a product it is complete: serverless public endpoints, dedicated isolated GPU endpoints, post-training and fine-tuning, and Data Lab dataset management all live together. For teams that would rather not maintain an inference platform, a fine-tuning platform and a dataset tool separately, that bundling is a real convenience. Each model also lists its deployment region (uk-south1 / us-north1) and the platform claims zero-retention data flow plus predictable latency, cost and data control — more useful for compliance-sensitive buyers than platforms that never say where a model runs.
Risks and limits
Three things must be accepted. First, there are no overseas closed models — it cannot help if you need Claude, GPT or Gemini. Second, the full price list and model catalogue are behind login, and even the visible prices are labelled approximate, so nothing can be verified from outside. Third, the product has been renamed and its paths migrated (AI Studio to Token Factory, with the old /prices-ai-studio now a 404) — a maintenance-risk signal worth weighing if you plan to depend on the API long-term.
Information verified 2026-10-15. Prices, throughput figures and product shape reflect what the nebius.com/ai-studio marketing page and the tokenfactory.nebius.com console homepage displayed on that date.
Related reviews
- Baseten: an inference cloud for open-weight models with a published $-per-million-token sheet and per-minute dedicated GPUs
- DeepInfra: low-latency inference for open-source models, transparent and extremely low pricing, supports crypto payment, includes image generation models
- Lambda Labs: GPU cloud compute plus inference API, 60-70% cheaper than AWS/GCP, a top pick for AI research
- Together AI: large-scale open-source model compute aggregation, 100+ models via one API, suited to both research and enterprise use
Quick facts
| Pricing model | Billed per token with prices shown straight on the console homepage alongside measured throughput: DeepSeek-V4-Pro-0813 at $1.32/1M in and $3.96/1M out (37 tok/s, uk-south1); GLM-5.3-Flash at $0.15/$0.50 (349 tok/s, us-north1); DeepSeek V4.1 Flash at $0.30/$1.20 (139 tok/s); the page notes prices are approximate and depend on actual usage, and the full price list lives behind login; the underlying GPU cloud bills per GPU-hour (HGX B200 $8.50, HGX H100 $4.50 and so on) |
|---|---|
| Model coverage | Featured models on the console homepage are open-weight models: DeepSeek V4 Pro, DeepSeek V4.1 Flash, Z.ai GLM-5.3-Flash, MiniMax M3 and Kimi K3; the full catalogue is behind login and no total model count is published |
| Latency / SLA | Measured throughput is published per model (GLM-5.3-Flash at 349 tok/s, DeepSeek V4 Pro at 37 tok/s); dedicated endpoints promise predictable latency, cost and data control plus zero-retention data flow; the company cites a Platinum rating in SemiAnalysis ClusterMAX 3.0, a third-party benchmark the vendor itself is quoting |
| Mainland direct connect | Proxy required (EU and US regions; the site makes no China-related statements) |
| Best for | Developers / Enterprise |
| Referral program | No public affiliate or rebate programme found. |
Pros
- Price and throughput are published together (for example GLM-5.3-Flash at $0.15/$0.50 with a stated 349 tok/s), which is far more useful than a bare unit price
- Full-stack: inference, fine-tuning/post-training, datasets and dedicated GPU endpoints in one place, so you do not have to stitch tools together
- Each model lists its deployment region (uk-south1 / us-north1) and the platform claims zero-retention data flow, which matters for compliance-sensitive buyers
Cons
- Open-weight models only — no GPT, Claude or Gemini, so it cannot solve access to closed frontier models
- The full price list and model catalogue sit behind login, and the visible prices carry an approximate disclaimer, so a complete cost estimate is not possible from outside
- The product has been renamed and migrated (AI Studio to Token Factory, with legacy paths such as /prices-ai-studio now 404), which is a maintenance-risk signal for anything built on its API long-term
Compare more AI API relays
See the full comparison board — filter by price tier, model coverage, and mainland direct-connect status.
Back to the comparison board →