Novita AI
Developer ToolsUpdated November 12, 2026
Developer cloud with one API for 200+ open-weight AI models, plus GPU instances, dedicated endpoints and isolated sandboxes for AI agents.
We may earn a commission when you sign up through our links. It does not affect our reviews.
What is Novita AI?
Novita AI is a developer cloud that puts three products behind one account and one API key: serverless APIs for more than 200 open-weight models, GPU capacity in the form of instances and dedicated endpoints, and an Agent Sandbox where AI agents run code, browsers and desktops in isolated environments. The LLM API follows the OpenAI format and a second endpoint accepts Anthropic SDK requests, so apps and tools like Claude Code or LangChain can switch to DeepSeek, Kimi, GLM or Qwen models by changing a base URL. Novita first became known for cheap Stable Diffusion APIs, but image generation is now a small corner of the catalog. Its strengths are low per-token prices and fast support for new open models. The weak spot is uneven quality across so many hosted models. Pick it if you build on open models and want inference, GPUs and agent runtimes on one bill.
How to start using Novita AI?
Create an account at novita.ai with Google, GitHub or an email address, then open Key Management in the console and click Create API Key. Give the key a name such as production or local-testing. The full key, which starts with
sk_, is shown only once, so copy it straight into a secrets manager or an environment variable calledNOVITA_API_KEY. Each account can hold up to 10 keys, and you can restrict every key to certain models or source IP addresses.Next, add credit. Novita is prepaid: payments go through Stripe by card (plus Apple Pay or Link in some regions), a manual top-up has to be more than $10, and cryptocurrency is not accepted. New accounts can receive a new-user voucher, and new Agent Sandbox users get $100 in promotional sandbox credits after a short setup survey, without a card. Turn on Automatic Top-Up and a low balance alert early, because an empty balance blocks API calls and terminates running sandboxes.
Your first request goes to
https://api.novita.ai/openai/v1/chat/completions. The quickstart uses thedeepseek/deepseek-v4.1-flashmodel, and any OpenAI SDK works once you change the base URL. If your code or agent speaks the Anthropic format, pointANTHROPIC_BASE_URLathttps://api.novita.ai/anthropicinstead; the docs include a full Claude Code setup. Before writing code, the LLM Playground lets you test a system prompt, a JSON schema and sampling settings such as temperature, top_p and repetition penalty.A successful call returns HTTP 200 with a non-empty
choices[0].message.content. From there, pick the next layer only when you need it. Dedicated Endpoints (Models Console, then Deployments) put a catalog or Hugging Face model on its own GPU with autoscaling, scale-to-zero and LoRA adapters. GPU instances start from the Explore page: choose a template or your own Docker image, a GPU and click Deploy. For agents, install the SDK withpip install novita-sandboxand create a sandbox from a template such as base or codex. Check spending under Billing, where usage can be exported per product.
Who is it for?
AI app developers
You ship chat, search or summarization features and closed-model bills keep growing. Novita gives you DeepSeek, GLM, Kimi, Qwen and Llama models through the same OpenAI-style client, with small models such as Llama 3.1 8B Instruct priced at $0.02 per million input tokens.
Coding-agent users
You live in Claude Code, Cursor or OpenCode and burn tokens all day. The Anthropic-compatible endpoint and the Coding Plan, which starts at $19.9 a month for 50M tokens, let you run open coding models at a fixed monthly cost instead of pure pay-per-token.
Agent builders
Your agent needs to run code, click through websites or operate a desktop without touching your own machines. Agent Sandbox starts isolated environments in under 200ms on average, can pause and resume with memory intact, and supports browser use and computer use with a live VNC view.
ML engineers who need GPUs
You want to fine-tune, render or serve a custom model without a long cloud contract. On-demand RTX 4090 instances list at $0.33 per hour, spot capacity is about half price, and serverless GPUs scale to zero when idle.
Privacy-minded teams
Legal will not approve a provider that trains on customer prompts. Novita’s Trust Center states that API inputs and outputs are not used for training by default, and lists a SOC 2 Type II audit and GDPR compliance.
Data and evaluation teams
You need to classify a large dataset or rerun a benchmark overnight. The OpenAI-compatible Batch API takes JSONL files, returns results within 24 hours and is currently offered at an introductory 50% discount on supported models.
Novita AI pricing
Novita has no free subscription tier. You top up a prepaid balance and pay per token for model APIs, per second for GPUs and sandboxes, and per GB for storage. The only subscription is the optional Coding Plan for coding agents. Enterprise buyers can negotiate higher rate limits, dedicated deployments and custom contracts. All prices below are in USD as listed on the official pricing pages and docs on 12 November 2026; model prices change often, so check the live table before budgeting.
GPT OSS 120B $0.05 input and $0.25 output, GLM 5.3 Flash $0.15 / $0.03 / $0.5, DeepSeek V4.1 Flash $0.3 / $0.006 / $1.2, GLM 5.3 $1.4 / $0.26 / $4.4, Kimi K3.1 Pro $2 / $0.2 / $10, Kimi K3 $3 / $0.3 / $15. Apodex 1.1 Mini is listed as free. Batch inference gets an introductory 50% discount on supported models.
Built-in web search costs $0.007 per Exa neural search request and $0.008 per Tavily basic search request.
MiniMax H3 video output costs $0.08 per second at 768P and $0.13 per second at 2K. Text-to-speech is $15 per 1M characters for Fish Audio S2 Pro and $25 per 1M characters for Inworld TTS 2.
50M tokens and up to 45 requests per minute, aimed at individual developers.
150M tokens and up to 150 requests per minute for small teams.
750M tokens and up to 450 requests per minute for larger teams or production use. Plan quota is used before your API balance, and auto-renewal can be cancelled at any time.
per GPU-hour, billed per second on running replicas only. RTX 4090 $0.61, RTX 5090 $0.73, H100 $1.99, H200 $2.99.
on-demand RTX 4090 $0.33/hr (spot $0.17/hr), L40S $0.55/hr, RTX 5090 $0.56/hr, RTX 6000 Ada $0.77/hr (spot $0.39/hr) and H100 SXM $3.39/hr (spot $1.70/hr); eight H100s cost $27.12/hr. Container disk includes 60 GB free, then $0.005/GB/day; volume disk is $0.005/GB/day and network volume $0.002/GB/day. Billing starts once the instance reaches the pulling state.
Each sandbox gets 20 GB of free ephemeral storage, and paused sandboxes use persistent storage with 60 GB free per account, then $0.00009/GB/h.
Custom. Higher rate limits, dedicated model deployments and contracts through the sales team.
max_tokens, because requests that the client abandons mid-stream (HTTP 499) are still billed for the tokens generated.Novita AI key features
One API for 200+ models
The catalog covers LLMs, vision, embeddings, rerankers, text-to-speech, video and a few image models, all called with one key. Chat models use the OpenAI request format, and a subset also works through the Anthropic SDK, which is how Claude Code and similar tools run on Novita models. Novita is also listed as an inference provider on Hugging Face and OpenRouter, so you can reach it through those routers too.
Prompt caching and reasoning controls
Cached input tokens cost roughly a tenth of the normal input price on supported models, which matters for agents that resend long histories. For hybrid reasoning models you can set enable_thinking to false to skip billed reasoning tokens, or ask for the reasoning in a separate field. You pay only for the thinking you actually need.
Dedicated Endpoints
Deploy a model from the Novita catalog or from Hugging Face on GPUs that no other customer shares, with autoscaling, scale-to-zero and LoRA adapter support. Pure text models keep the OpenAI-compatible chat API, so the switch from serverless is mostly a change of model name. This is the route when shared serverless limits or noisy neighbours start to hurt latency.
GPU Cloud
GPU instances give you full-control machines from public templates (PyTorch, JAX, CUDA) or your own Docker image, with spot pricing at about half the on-demand rate and managed network volumes for shared data. Serverless GPUs run your container as workers that scale up under load and back to zero, billed per second of running time.
Agent Sandbox
Isolated Linux sandboxes start in under 200ms on average and can run Python, JavaScript and shell commands, browser automation or a full desktop for computer use with a live VNC view. Pausing preserves the filesystem, memory and running processes, and stops CPU and RAM billing. Free accounts get 5 concurrent sandboxes, paid accounts 100 and enterprise accounts 500 or more.
Spend and access controls
Team accounts can set fixed or monthly budgets per member and per API key, and each key can be limited to chosen models or IP addresses. Low balance alerts, automatic top-up and per-product bill exports help finance teams see where the money goes. Rate limits rise through tiers T1 to T5 as your monthly top-ups grow.
Novita AI Pros & Cons
Pros
Novita covers the whole open-model stack in one place: per-token APIs, dedicated GPU endpoints, raw GPU instances and agent sandboxes share a single balance and API key. Per-token prices on popular open models are competitive, cached input is cheap, and new releases such as GLM, DeepSeek and Kimi versions tend to appear quickly. Drop-in OpenAI and Anthropic compatibility means existing code and coding agents work after a base URL change. The no-training policy, SOC 2 Type II audit and per-key budgets make it easier to get past a security review than with many low-cost hosts. Start by pinning two or three well-used models and setting a monthly budget per key.
Cons
Quality is not uniform across hundreds of models. Developers report errors on older or less popular models, a model served without tool calling enabled, and occasional output differences compared with the original lab’s API, so you should run your own evals before switching production traffic. Refunds on topped-up credit are generally not available, and client-side disconnects are still billed. Rate limits are tied to how much you top up, which slows down new accounts with bursty traffic. If you came for image generation, the catalog now offers very few image models. Keep a fallback provider configured and test the exact model ID you plan to use.
Use Novita AI if you want…
- Call open-weight LLMs through an OpenAI-compatible API.
- Run Claude Code or Cursor on cheaper open models.
- Cut token costs with prompt caching and batch jobs.
- Deploy a Hugging Face model on a dedicated GPU.
- Rent RTX 4090 or H100 GPUs by the second.
- Give AI agents isolated sandboxes for code and browsers.
- Keep prompts out of model training by default.
- Pay a fixed monthly price for coding-agent tokens.
Novita AI alternatives
Fireworks AIRead review →
OpenRouterRead review →
ReplicateRead review →
falRead review →
RunPodRead review →
Vast.aiRead review →
ModalRead review →Novita AI Reviews
Developers who route traffic through OpenRouter often name Novita among the few providers that did not degrade model quality, and some prefer it for zero data retention and good prompt-cache hit rates.
Users building on DeepSeek, GLM and Kimi models like the aggressive pricing and data retention terms for production work, though they note that discounts come and go as hosts undercut each other on popular models.
Critics say Novita hosts so many models that some get too little testing, pointing to frequent errors on an older Kimi release, a Gemma endpoint without tool calling and suspected quantization losses against other providers.
One developer who tried several inference services, Novita included, said inconsistent performance and pricing left them unsatisfied, and they now host their models locally instead.
Early Product Hunt reviews, written when Novita mainly sold image APIs, praise it as a cheap and stable Stable Diffusion API, but they predate today’s LLM, GPU and sandbox platform.
FAQ
Does Novita AI have a free plan?
There is no free subscription tier; you pay as you go from a prepaid balance. New users can receive a new-user voucher, new Agent Sandbox users can claim $100 in sandbox credits after a setup survey, and a few models such as Apodex 1.1 Mini are listed as free.
Does Novita AI use my prompts to train models?
According to its Trust Center, Novita does not use API prompts, inputs, outputs or conversation histories to train or improve models by default. The company states it is SOC 2 Type II audited and GDPR compliant.
Is the Novita AI API compatible with the OpenAI and Anthropic SDKs?
Yes. Use https://api.novita.ai/openai as the base URL for OpenAI SDKs and https://api.novita.ai/anthropic for the Anthropic SDK, which also lets Claude Code run on supported Novita models.
What are the rate limits on Novita AI?
LLM limits depend on your account tier, from T1 to T5, based on monthly top-ups over the last three calendar months. You can request higher limits through Discord or the sales team, and sandbox concurrency is 5 on free accounts and 100 on paid accounts.
Can I still generate images with Novita AI?
Only in a limited way. The current catalog lists the ming image 0.1 design text-to-image and layer models plus MiniMax H3 for video, so for a wide choice of image models you would look at fal or Replicate.
Can I get a refund on unused credit?
Refunds are generally not available. In exceptional cases such as complete service unavailability, support can review a request, and Coding Plan subscribers can cancel auto-renewal at any time.
What payment methods does Novita AI accept?
All payments run through Stripe, so you can pay by credit or debit card and, in some regions, wallets like Apple Pay or Link. Cryptocurrency is not supported, and manual top-ups must be more than $10.
Where do Novita AI sandboxes run?
Agent Sandbox currently runs in two US regions: us-phx-1, the default, and the legacy us-virginia-1 region. Newer features such as secrets and snapshots are available only in us-phx-1.
Similar tools
You.comDeveloper platform of web search, content extraction, cited research and model inference APIs that feed AI agents and RAG apps real-time web
QodoAI code review platform that runs specialized agents on every pull request, enforces team rules and checks changes against tickets and linked
Applitools EyesVisual AI testing tool that adds screenshot checkpoints to Playwright, Cypress or Selenium tests and flags real UI regressions across browsers.