Novita AI
Novita AI is an AI-native cloud for developers: over 200 open models behind a single API, isolated agent sandboxes, and on-demand GPUs. Everything is billed by usage, with 100 USD in starting credits.
What is Novita AI?
Novita AI is an AI-native cloud built around three services that share one account, one API and one usage-based bill: serverless model APIs, an agent sandbox, and GPU infrastructure.
The model layer exposes more than 200 open models through a single endpoint. Language models from DeepSeek, Qwen, GLM, Kimi, MiniMax, Gemma, ERNIE, Ling, Nemotron and StepFun sit alongside image, audio and video generators, embedding models and an AI Search building block. Everything is billed per token rather than per hour, and the homepage advertises 200 ms latency and 99.5% uptime. Teams that need predictable throughput can move a model onto a dedicated endpoint with isolated resources.
The Agent Sandbox is a runtime rather than a notebook. Each task runs in a system-isolated environment that starts in under 200 milliseconds on average, and thousands can run in parallel. Inside it, an agent executes Python, JavaScript or C++, calls external APIs, drives a browser, controls a full desktop through computer use, pauses and resumes long sessions, and streams its screen over VNC. Novita positions this for coding agents, browser automation, evaluations and reinforcement learning, AI-assisted CI/CD pipelines and long-running workflows.
The GPU layer comes in three levels: full-control GPU instances, serverless GPUs that scale to zero in 30 seconds, and dedicated bare-metal clusters. H100 and H200 nodes are advertised with fourth-generation NVLink and 400 Gb/s RDMA networking, spread across 14 regions and more than 20 locations. Provisioning goes through REST, gRPC, Terraform or a CLI, with ready-made PyTorch, JAX and CUDA templates.
The economic argument is repeated throughout the site: up to 50% below major cloud providers, spot capacity, per-second billing and nothing charged while idle. Around the products sits a complete developer ecosystem, with versioned documentation, an API reference, a dated changelog, SDKs, an llms.txt index, an installable agent skill on GitHub, a Discord community and a public status page. Named customers include Hugging Face, Quora's Poe, OpenRouter, TiDB, Wiz and Fish Audio, and Hugging Face's co-founder and CTO is quoted on how quickly new models go live.
What it does
- Call more than 200 text, image, audio, video, vision and embedding models through one API
- Run AI-generated code in isolated sandboxes that start in under 200 milliseconds
- Let agents drive a browser or a full desktop environment to complete multi-step tasks
- Rent GPU instances on demand or as spot capacity, or submit jobs to serverless GPUs
- Deploy a private dedicated endpoint with reserved resources and stable latency
- Fine-tune, train from scratch or run reinforcement learning on H100 and H200 clusters
- Automate the whole platform from code through REST, gRPC, Terraform or the CLI
When to use Novita AI / When not to
A quick filter to help you decide if Novita AI is the right fit.
When to use Novita AI
- Backend and full-stack developers who want to call many open models through one API instead of wiring up a provider for each
- Teams building autonomous agents that need an isolated runtime for code, browser and desktop actions
- ML and MLOps engineers running inference, fine-tuning or reinforcement learning on rented H100 and H200 capacity
- Cost-sensitive startups that would rather pay per token and per second than commit to a large cloud contract
- Researchers and graduate students who need short bursts of GPU time without a procurement process
When not to use Novita AI
- Non-technical users looking for a ready-made chatbot or a no-code interface: everything here is code and API keys
- Organisations that need a contractual guarantee on where data physically sits, since no hosting region is ever named
- Regulated healthcare, federal or financial workloads, as the service is not configured for HIPAA, FISMA or GLBA by default
- Buyers who require a published DPA, a named subprocessor list and an identified legal entity before signing
- Anyone located in a country under US embargo or listed on a restricted-party list, who is contractually excluded
How to use Novita AI
A typical end-to-end flow, from setup to results.
- Create an account from the registration page, using an email address or a Google or GitHub identity
- Claim the 100 USD of starting credits, which require no payment card and stay valid for 90 days
- Open the web console to generate an API key and follow the Quickstart guide
- Call a model with a standard HTTP request, or browse the catalogue to compare per-token prices first
- For agents, install the sandbox SDK or CLI and launch a first sandbox from a reusable template
- Add a browser or desktop session to the sandbox when the agent needs to act outside pure code
- For heavier work, pick a GPU instance in the console, or provision it from code with the API or Terraform
- Move a production workload onto a dedicated endpoint or serverless GPU once traffic becomes predictable
- Top up the balance by card through Stripe or by PayPal, and set budgets, automatic top-up and low-balance alerts
- Monitor spend and LLM API metrics in the console, and reach the team by email, Discord or a sales meeting
Pros & Cons
Pros
- Very broad catalogue of open models behind one API, with new releases reportedly supported on day one
- Per-model prices published openly, readable without creating an account
- Genuine usage-based billing: per token, per vCPU-second and per GPU-second, with no mandatory subscription
- 100 USD of free credits granted without a payment card
- Contractual Zero Data Retention policy and an explicit commitment not to train on customer content
- Three levels of abstraction on one platform, from a simple API call to raw bare-metal hardware
- Complete developer ecosystem, verifiable customer references, an AICPA SOC 2 badge and a public status page
Cons
- No full legal entity name and no postal address anywhere on the site
- GDPR compliance is never claimed, and no DPA or named subprocessor list is published
- No data residency commitment: the 14 advertised regions are never identified
- All fees are non-refundable in principle, and liability is capped at six months of spend or 500 USD
- GPU rates and Coding Plan tiers are not readable without JavaScript, which makes comparison harder
- A single generic email address handles support, legal and privacy requests alike
- Pay-as-you-go instances may fail to restart after being stopped if their resources have been reclaimed
Pricing & Plans
A free entry point is available: new accounts receive 100 USD of credits, valid for 90 days and granted without a payment card, and the agent sandbox keeps a permanent free tier of five concurrent sandboxes limited to one-hour sessions. Several models are also billed at zero. Beyond that, pricing is strictly consumption-based rather than plan-based, so no monthly entry price applies. The lowest published rates observed are 0.14 USD per million input tokens on the cheapest language models, 0.0000098 USD per vCPU-second and 0.0000032 USD per GiB-second in the sandbox. GPU rates and the tiers of the monthly Novita Coding subscription are not published in readable form.
- 100 USD of credits
- 5 concurrent sandboxes
- 1-hour maximum session
- 2 vCPU and 4 GB RAM per sandbox
- no priority scheduling
- pay-as-you-go
- 100 concurrent sandboxes
- sessions up to 24 hours
- up to 8 vCPU and 8 GB RAM
- priority scheduling
- unlocked automatically once the balance is above zero
- 99.95% uptime SLA
- custom vCPU ceiling
- deployment region of your choice
- unlimited concurrent instances
- arranged by email
- pay per million input and output tokens
- with a separate cache rate and a 50% introductory discount on batch inference
- private endpoints with isolated resources for guaranteed performance
- on-demand instances
- spot capacity advertised at up to 50% less
- serverless GPUs billed per second
- bare-metal clusters
- and monthly subscription instances with pre-reserved resources
- monthly subscription covering nine models
- upgradable to higher tiers and cancellable at any time
- with amounts not published in readable form
- Custom enterprise plans negotiated through an order form or a master service agreement
Data, GDPR & hosting
A consolidated view of how Novita AI handles your data.
GDPR overview
Novita AI never claims GDPR compliance anywhere on its site. The single occurrence of the word GDPR places the obligation on the customer, who must ensure that any third-party personal data in their inputs is lawfully processed. The privacy policy, in force since 13 May 2026, is built around Californian law: CCPA and CPRA categories, a Do Not Sell or Share mechanism, and a 45-day response window. European and UK residents are nonetheless granted access, deletion, rectification, portability, consent withdrawal and objection rights, and international transfers are said to rely on Standard Contractual Clauses and, where applicable, the EU-U.S. Data Privacy Framework. No Article 27 representative and no data protection officer are named, no DPA is published, and requests all go to a single generic support address.
Who owns the data?
You keep ownership of what you send. The terms state that you retain copyright and any other proprietary rights in your Input, and Novita AI claims no ownership over your contributions. In exchange you grant the company a worldwide, royalty-free, non-exclusive licence to reproduce, view and use both your Input and the generated Output, but strictly for the purpose of delivering the service to you. One exception deserves attention: suggestions and feedback you send about the platform are treated as Submissions and become the exclusive property of Novita AI, which may use and disseminate them for any lawful purpose without compensation.
Reuse rights
Nothing in the terms restricts what you do with the output you generate: Novita AI describes itself as a service provider, keeps no ownership claim over your content, and grants its own licence only for the time needed to run the service. You therefore reuse your inputs and outputs freely, without asking permission, subject to your own upstream rights and to the acceptable use policy. The company also commits, by default, not to train its models on your content and not to log it for human review, although automated safety screening is still applied. Practical limits come from elsewhere: you remain responsible for holding the rights to whatever you submit, for the lawfulness of any third-party personal data it contains, and for export control and sanctions rules.
Data retention & training
Hosting summary
Novita AI publishes no hosting jurisdiction. The GPU pages advertise 14 global regions, more than 20 locations and four or more continents, but none of them is named, and only the enterprise sandbox tier mentions a deployment region of your choice. The privacy policy states that personal information may be transferred to countries outside your own, expressly including the United States, and that Standard Contractual Clauses, the EU-U.S. Data Privacy Framework where applicable and other recognised mechanisms are used when the law requires them. Cloud storage providers appear in the list of third-party categories, but none is named. On the technical side, the domain resolves to an anycast address geolocated in the United States, which describes the website rather than any data residency commitment. A Vanta-hosted trust centre exists but its content cannot be read without JavaScript, so no certification scope or subprocessor list could be verified from the published pages.
Where Novita AI works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Novita AI.
- The operating entity is opaque: only the brand name appears, with no company registration details and no postal address
- GDPR compliance is never claimed, no data processing agreement is published and no subprocessor is named
- No hosting region is identified, so nothing prevents data from being processed in a jurisdiction you did not expect
- Costs can escalate quietly: consumption billing per token and per second rewards experimentation and punishes a runaway agent
- Fees are non-refundable by default and liability is capped at six months of spend or 500 USD, whichever is lower
- Agents given browser and desktop control act on real systems, so an unsupervised loop can cause real-world consequences
- Model outputs are explicitly disclaimed as possibly inaccurate or biased, and verifying them remains entirely on you
Setup & Integrations
Technical difficulty
Aimed squarely at developers: there is no no-code path, and every route starts with an API key. The lightest entry is a single HTTP call documented in the quickstart, which any backend developer can complete in minutes. The sandbox requires installing an SDK or CLI, though reusable templates remove most configuration. The GPU layer is the demanding part, assuming familiarity with instances, container and network volumes, CUDA versions and images. Signing up with Google or GitHub and free credits without a card lower the initial barrier considerably.
Deployment
Integrations
Supported languages
Behind Novita AI
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
How many models does Novita AI serve, and of what kind?
Is there anything free?
How is the service billed?
Is customer data used to train models?
Where is the data hosted?
Which payment methods are accepted?
Can charges be refunded?
Is there an API and proper documentation?
What is the minimum age to use the service?
Is there an affiliate programme?
Should you pick Novita AI?
Novita AI is a serious piece of infrastructure rather than a finished product, and it should be judged as such. Everything a developer needs in order to decide is published: a catalogue of more than 200 open models with per-token prices visible without an account, three levels of GPU access, an agent runtime that goes well beyond a hosted notebook, and tooling that ranges from a CLI to Terraform. The commercial promise is consistent throughout, and the customer references, the SOC 2 badge and the public status page all support it.
Two reservations matter. The first is corporate transparency: the site names no legal entity beyond the brand and publishes no postal address, so the only clue to jurisdiction is a Delaware governing-law clause. The second is data governance. GDPR compliance is never claimed, no data processing agreement is available, no subprocessor is named, and none of the 14 advertised regions is identified. That combination is manageable for a startup shipping a prototype, and much less so for a European organisation with a procurement checklist.
The contractual commitments on customer content pull in the opposite direction and deserve credit: a default Zero Data Retention clause and an explicit statement that personal information is not used for model training are stronger than what many competitors offer. Anyone signing up should still note that all fees are non-refundable by default, that liability is capped at six months of spend or 500 USD, and that GPU rates cannot be read without JavaScript.
In short, a strong technical and economic fit for builders who want breadth of models and elastic compute at low cost, and a platform that a compliance-driven buyer should question directly before committing.
- Choosing a selection results in a full page refresh.
- Opens in a new window.