Buy If / Skip If
Quick decision guide. PromptQuorum assessment: a good fit for API-first teams, a weaker fit for anyone who needs published per-GPU rates before signing up.
- Use Novita AI if:
- You want open-weight models behind an OpenAI-style API, billed per token
- You already use LlamaIndex and want a documented Novita integration
- You need an occasional GPU instance or serverless endpoint under the same account
- You prefer per-second GPU billing over rounding up to the hour
- Skip Novita AI if:
- You must pay in cryptocurrency (the billing FAQ says it is not supported)
- You need a refund-friendly prepaid model (the FAQ says refunds are generally not available)
- You need the cheapest raw GPU hour: compare a marketplace in our RunPod review first
- You need guaranteed data-residency or retention terms that you have verified in writing
Novita AI Scorecard
PromptQuorum assessment, based on official Novita AI pages and documentation checked 2026-11-22. This is not a hands-on lab test. Scores reflect fit for developers who want hosted open models plus optional GPUs.
Price
- Score:
- 8.5/10
- Why:
- Low per-token rates listed; GPU hourly rates not readable
Reliability
- Score:
- Not rated
- Why:
- No uptime figure verified; rated on stated design only
Ease of setup
- Score:
- 8.5/10
- Why:
- API key plus model name; documented LlamaIndex steps
GPU choice
- Score:
- 7/10
- Why:
- GPU types and rates not readable on the pages checked
Support and docs
- Score:
- 8/10
- Why:
- Docs cover billing, instances and integrations
Overall
- Score:
- 7.8/10
- Why:
- Strong API product; verify GPU rates in the console
Every score is PromptQuorum assessment derived from official pages (see "How We Evaluate"), not from PromptQuorum running workloads on the platform.
Verdict: Who Should Use It
Use Novita AI if the model API is what you want: open-weight models at low per-token rates with an OpenAI-style interface. On the pricing pages checked 2026-11-22, GPT OSS 120B is listed at $0.05 / $0.25 per 1M input / output tokens, against $0.15 / $0.60 on Together AI and Fireworks (the Together AI input rate is labelled "(cached)").
Do not choose it for GPU work on price alone, because the official pages checked did not expose per-GPU hourly rates. Open the console, read the live rate for the exact card, and compare it with the numbers in cloud GPU cost per hour before you commit. The no-refund default makes small test top-ups the sensible first step.
How Novita AI Works
One account, three products: a model API billed per token, GPU instances billed per second, and serverless GPUs billed per second of use. Descriptions below come from Novita AI's own pages.
- Model API: a catalog of open and partner models (Llama, DeepSeek, Qwen, Gemma, GLM, Kimi, GPT OSS) with per-token prices, plus embeddings and search APIs listed per request.
- GPU instances: you pick a GPU type, region and image, and billing starts when the instance reaches the Pulling status and stops when you stop it.
- Serverless GPUs: Novita states per-second billing, billing only for consumed resources, scale to zero in 30 seconds and sub-second startup after pre-warming; it gives no cold-start figure for instances that are not pre-warmed.
- LlamaIndex: the official LlamaIndex guide has you copy a model name from the Model Library, create an API key in platform settings and use the
NovitaAIclass; thellama-index-llms-novitapackage is listed on PyPI. - Batch inference: the pricing page describes an introductory 50% discount on input and output tokens for supported models.
Pricing and Billing
Model API rates are public and low; GPU hourly rates were not readable on the official pages checked, so none are quoted here. All figures are USD per 1M tokens from the official pricing page, checked 2026-11-22.
- GPU instance billing (docs, checked 2026-11-22): unit price x duration x number of cards, accurate to the second and settled hourly; see the GPU instance pricing docs.
- Storage (same docs): first 60 GB of container disk free, then $0.005 per GB per day; volume disk $0.005 per GB per day; network volume $0.002 per GB per day.
- GPU hourly rates: not verified. The live table is on the GPU pricing page and in the console; Novita also notes a GPU pricing adjustment effective 2026-03-16, so older figures elsewhere may be out of date.
- Serverless rates: not verified; Novita states spot pricing can save "up to 50%" without listing rates in the page text.
- Payments: processed by Stripe; per the billing FAQ, cryptocurrency is not supported. Some regions may offer local wallets.
- Pricing changes often: treat every number as a snapshot and confirm on the live page.
OpenAI GPT OSS 120B
- Input / output:
- $0.05 / $0.25
- Context:
- 131K
Llama 3.1 8B Instruct
- Input / output:
- $0.02 / $0.05
- Context:
- 16K
Llama 3.3 70B Instruct
- Input / output:
- $0.135 / $0.4
- Context:
- 12K
Gemma 3 27B
- Input / output:
- $0.119 / $0.2
- Context:
- 98K
Qwen3.5 27B
- Input / output:
- $0.3 / $2.4
- Context:
- 262K
DeepSeek R1 0528
- Input / output:
- $0.7 / $2.5
- Context:
- 164K
Model API vs Together AI and Fireworks
For the one model all three list, GPT OSS 120B, Novita AI shows the lowest per-token price on the pages checked 2026-11-22 (the Together AI input rate is labelled "(cached)"). The comparison covers list prices only, not latency, quality or limits, which this review did not measure.
- Together AI lists its GPT OSS 120B input rate as "(cached)", so confirm the uncached input rate on the model page.
- The H100 figures are different products: Together AI lists dedicated inference, Fireworks lists on-demand deployments. They are not a like-for-like quote.
- Fireworks states region-restricted deployments are priced at 1.5x.
Novita AI
- GPT OSS 120B in / out:
- $0.05 / $0.25
- H100 per hour:
- Not verified
Together AI
- GPT OSS 120B in / out:
- $0.15 / $0.60
- H100 per hour:
- $5.49 dedicated
Fireworks
- GPT OSS 120B in / out:
- $0.15 / $0.60
- H100 per hour:
- $8.00 on-demand
Fit for Local LLM Work
Novita AI is a way to run models you could not fit on your own card, without renting a whole GPU. Our VRAM cheatsheet for DeepSeek-R1 distills puts the 32B distill at about 20.5 GB and the 70B at about 44 GB at Q4_K_M, which is why large models often go to an API or a multi-GPU rental.
- Model fits your card: run it locally; see how much VRAM a local LLM needs.
- Model does not fit: call it through the Novita model API per token instead of renting a pod by the hour.
- Custom weights or code: use a GPU instance or serverless endpoint and size the card from the VRAM figures above.
- No tokens-per-second figures are quoted: Novita does not publish a verified throughput table on the pages checked, and PromptQuorum did not benchmark it.
Use Cases
Matches between Novita AI products and common jobs, from strong to weak fit (PromptQuorum assessment).
- Strong fit: prototyping an app against open-weight models with an OpenAI-style client and per-token billing.
- Strong fit: RAG pipelines in LlamaIndex using the documented
NovitaAIclass and an embedding model such as BGE-M3 ($0.01 per 1M tokens). - Good fit: bursty inference on your own model through serverless GPUs, where you pay only while requests run.
- Weaker fit: long-running training that needs a published, fixed hourly rate; confirm the live GPU rate first, or see our vLLM explainer for self-hosted serving options.
- Poor fit: workloads that need crypto payment or prepaid money you can get back.
Risk and Trust
Facts from Novita AI's own documentation, checked 2026-11-22. Unverified items are marked as such.
- Billing model: USD, pay as you go. GPU compute is billed to the second and settled hourly; the model API is billed per token.
- Payments: all payments are processed by Stripe; cards are accepted and the FAQ states cryptocurrency payments are not supported.
- Refunds: the FAQ says "Refunds are generally not available". In exceptional cases, such as complete service unavailability, you can request one from support.
- Credits: sign-up credit vouchers can be revoked when suspicious registration activity is detected, such as multiple accounts from one device.
- Hardware source: Novita describes its own GPU cloud with global regions. The region count differs between its pages, so no number is quoted here.
- Data handling: no statement on prompt storage or retention was readable in the pages checked: not verified. Read the privacy policy and terms before sending sensitive data.
- Account balance: the FAQ states the Coding Plan requires a positive account balance. Other arrears and shutdown rules were not found: not verified.
Pros and Cons
A low-cost API with optional GPUs; the gaps are around published GPU rates and refunds.
- Pros
- GPT OSS 120B listed at $0.05 / $0.25 per 1M tokens
- Per-second GPU billing, no hourly rounding
- Documented LlamaIndex integration
- Batch inference discount described on the pricing page
- Cons
- GPU hourly rates not readable on the pages checked
- Refunds generally not available
- No cryptocurrency payments
- No verified data-retention statement found
Where to Sign Up
Start with a small top-up and the model API, then add GPUs once you have read the live rate. Links go to Novita AI's official pages.
- 1Create an account on Novita AI and open the console.
- 2Create an API key in platform settings.
- 3Pick a model in the Model Library and copy its model name.
- 4Make a small Stripe top-up and run a test call.
- 5For GPU work, read the live hourly rate for your exact card before launching an instance.
Alternatives to Consider
The right alternative depends on whether you want an API, a rented pod or your own hardware.
- Rented pods with published rates → RunPod review.
- Peer-to-peer GPU marketplace → Vast.ai review.
- Other hosted APIs → Together AI and Fireworks; see also best LLM API gateways to route between providers.
- API versus local → local LLMs vs cloud APIs.
- Fine-tuning on rented GPUs → best LLM fine-tuning frameworks.
- Run it yourself first → how to install Ollama, and weigh the build versus rent calculator, or compare RunPod vs Vast.ai pricing.
How We Evaluate
This review is not a hands-on test: PromptQuorum did not run workloads on Novita AI. It is built from official pages and documentation fetched on 2026-11-22.
- Official, checked 2026-11-22: pricing page, GPU instance pricing docs, billing FAQ, LlamaIndex guide, serverless page.
- Not verified: per-GPU hourly rates, serverless rates, uptime, throughput, data retention. The live GPU rate table could not be read from the pages fetched.
- PromptQuorum assessment: the scorecard, fit ratings and recommendations are editorial judgment on those sources.
- Comparison prices for Together AI and Fireworks come from their own pricing pages, checked the same day.
Frequently Asked Questions
What is Novita AI?
Novita AI is a developer platform with a per-token model API catalog, GPU instances billed per second and serverless GPUs. Payments are processed by Stripe. This answer reflects Novita AI's official pages checked 2026-11-22.
How much does the Novita AI model API cost?
On the official pricing page checked 2026-11-22, GPT OSS 120B is $0.05 per 1M input tokens and $0.25 per 1M output tokens, Llama 3.1 8B Instruct is $0.02 / $0.05, and Llama 3.3 70B Instruct is $0.135 / $0.4. Prices change, so confirm on the live page.
How are Novita AI GPU instances billed?
Compute is unit price x duration x number of cards, accurate to the second and settled hourly. Billing starts when the instance reaches the Pulling status and stops when you stop it. Storage beyond the first 60 GB of container disk costs $0.005 per GB per day.
What are the Novita AI GPU hourly prices?
Not verified. The official documentation points to the live GPU pricing page, whose rate table could not be read in this check. Open the console for the current rate of the exact GPU, region and product you want.
Does Novita AI offer serverless GPUs?
Yes. Novita states per-second billing, scale to zero in 30 seconds and sub-second startup after pre-warming. It gives no cold-start figure for instances that are not pre-warmed, and no serverless rates were readable in this check.
Does Novita AI work with LlamaIndex?
Yes. The official guide has you copy a model name from the Model Library, create an API key in platform settings, install LlamaIndex and use the NovitaAI class. The llama-index-llms-novita package is listed on PyPI.
Can I pay Novita AI with cryptocurrency?
No. The billing FAQ states that cryptocurrency payments are not supported. Payments are processed by Stripe with credit and debit cards, and some regions may offer local wallets.
Does Novita AI give refunds?
Refunds are described as generally not available. In exceptional cases such as complete service unavailability, you can request one from support. Start with a small top-up while you test.
Is Novita AI cheaper than Together AI and Fireworks?
For GPT OSS 120B, the pages checked 2026-11-22 list Novita AI at $0.05 / $0.25 per 1M tokens against $0.15 / $0.60 on both Together AI and Fireworks. This compares list prices only, not speed or quality. GPU rates were not compared.
Is my data private on Novita AI?
Not verified. No statement on prompt storage or retention was readable in the pages checked. Read the privacy policy and terms, and do not send sensitive data to any hosted API without confirming the terms in writing.
Is Novita AI a good choice for running local LLMs?
It is a good choice when a model does not fit your own card: call it per token instead of renting a pod. For a model that fits locally, run it locally. No throughput figures were verified, so this review quotes none.
