Novita AI is the most consistent runner-up in our weekly price checks: rarely the single cheapest host, but within cents of the leader across one of the widest open-model catalogs, and occasionally tied for first. If DeepInfra is the default cheap host, Novita is the default second source, and having a second source matters the day your primary throttles you.
Numbers from our latest verified check follow; the live rankings hold the current table per model.
Novita LLM pricing, per 1M output tokens
| Model | Novita | Table leader |
|---|---|---|
| Llama 3.1 8B | 5 cents | Tied for cheapest |
| gpt-oss-20b | 15 cents | DeepInfra, 14 cents |
| gpt-oss-120b | 25 cents | DeepInfra, 17 cents |
| Qwen2.5 72B | 40 cents | Tied for cheapest |
| Llama 4 Scout | 59 cents | DeepInfra, 30 cents |
| Qwen3 235B A22B | 80 cents | Hyperbolic, 40 cents |
| Llama 4 Maverick | 85 cents | DeepInfra, 60 cents |
| DeepSeek-V3 | 1.12 dollars | DeepInfra, 89 cents |
| MiniMax M2 | 1.20 dollars | OpenRouter, 1.02 dollars |
| Qwen3 Coder 480B | 1.30 dollars | DeepInfra, 1.00 dollar |
| GLM-4.6 | 2.20 dollars | Tied with official |
| DeepSeek-R1 | 2.50 dollars | DeepInfra, 2.15 dollars |
| Kimi K2 | 2.50 dollars | DeepInfra, 2.00 dollars |
The pattern: on small models Novita ties the leaders; on larger ones it typically sits 15 to 40 percent above DeepInfra but well below official endpoints and most Western hosts. Where it stands out is coverage: Chinese-family models (GLM, MiniMax, Kimi, Qwen variants) show up on Novita early and reliably.
The signup credit
Novita hands new accounts roughly $10 in referral or signup credit, enough to benchmark several models properly before paying. It is one of the standing offers in our catalog, alongside the bigger startup credit programs that apply once you have an entity.
When Novita is the right pick
- You want one key with wide open-model coverage, including new Chinese releases, without waiting for other hosts to add them.
- You need a second source for failover: OpenAI-compatible API, so pointing a fallback at Novita is trivial.
- You are optimizing for price but your exact model is one where Novita ties the lead (check the model's page from the rankings).
When it is not: if a single model dominates your bill and DeepInfra serves it 20 to 40 percent cheaper, go where the table points. This is a per-model decision, not a brand decision.
Bottom line
Novita in 2026 is competitively priced, unusually broad, and the natural failover behind whichever host tops your model's table. Take the current numbers from the live rankings, grab the signup credit from the catalog, and split traffic if reliability matters more than the last 15 percent of price.
Related reading: cheapest AI inference APIs, DeepInfra pricing, OpenRouter vs direct APIs.