Novita is an inference cloud platform from Novita AI. It gives you access to more than 200 AI models — chat, code, image, video, and speech — through a single API key and an OpenAI-compatible endpoint. Beyond model APIs, the same console covers on-demand GPU instances for running your own models, serverless GPUs billed by the second that scale to zero, and agent sandboxes that execute AI-generated code in isolation. There is no fixed monthly plan: you pay per token consumed or per hour of GPU runtime. The company is based in the United States.
Key Features
- 200+ models behind one API: Chat, code generation, embeddings, and reranking models can all be called with a single API key. The endpoint is OpenAI-compatible (
https://api.novita.ai/openai), so migrating usually means swapping the base URL and key in your existing OpenAI SDK code - More than text: Image generation, image editing, video generation, text-to-speech, voice cloning, and even search APIs are available from the same account, so multimodal features do not have to be spread across several vendors
- GPU cloud (on-demand and spot): Create, start, stop, and delete GPU instances from the API or console, then connect remotely to run your own models. Spot instances are offered at up to 50% off the on-demand rate
- Serverless GPUs and dedicated endpoints: Per-second billing, no concurrency cap, and scale-to-zero after about 30 seconds of idle time. Custom models can be deployed as dedicated endpoints without managing infrastructure yourself
- Agent sandbox: A way to run AI-generated code safely in an isolated environment. It supports command execution, SSH, filesystem operations, Git integration, and interactive terminals, plus templates, snapshots, and secret management
- Batch processing and an MCP server: A batch API handles large volumes of requests asynchronously. An official MCP server (novita-mcp-server) is published on GitHub, so MCP-capable AI clients can work with the platform directly
Pricing
There is no fixed monthly subscription — everything is usage-based per feature. The main billable items are as follows.
| Billable item | Approximate price |
|---|---|
| LLM API (per token) | Varies by model. Lightweight models start around $0.02 per million input tokens; larger models run to several dollars. Some models have a discounted cache-read rate |
| Image generation | $0.0225–$0.36 per image (Flux.1 family) |
| Video generation | $0.084–$0.168 per second, or $0.25–$2.25 per video depending on the model |
| Audio | Text-to-speech at $15 per million characters; voice cloning at $0.1–$1.5 per voice |
| Search API | $0.007–$0.016 per request, depending on search type |
| GPU instances | Hourly on-demand billing, with spot instances up to 50% cheaper. Rates differ per GPU type, so check the official pricing table |
| Serverless GPU | Per-second billing for consumed resources only, scaling to zero after roughly 30 seconds idle |
Pricing is as of August 2026. Check the official site for current rates.
Pros & Cons
✅ Pros
- The OpenAI-compatible API keeps migration changes to a minimum
- No fixed monthly fee, so a proof of concept can start for very little
- Model APIs and GPU instances live in the same account, making the “start with the API, move to dedicated GPUs once load is understood” path straightforward
- Image, video, speech, and search all land on one bill
- Spot instances and scale-to-zero serverless give you levers for controlling cost
⚠️ Cons
- This is a developer platform, not a ready-made chat UI for general users — API keys and code are assumed
- Pricing is split finely across models and GPU types, which makes an up-front total estimate hard to produce
- The model lineup turns over quickly, so implementations tightly coupled to specific model names age fast
- Compared with the major clouds, there is less Japanese-language material and fewer community case studies
- SLA, support terms, and data handling need to be verified individually for business use
Comparison with Similar Services
| Comparison point | Novita | Together AI | Fireworks AI | RunPod |
|---|---|---|---|---|
| Main scope | Model API + GPUs + sandbox | Model API + GPU clusters | Model API (fast inference focus) | GPU rental + serverless |
| OpenAI-compatible API | Yes | Yes | Yes | Depends on your deployment |
| Image / video / audio | Supported | Partial | Partial | Depends on the model you run |
| GPU instances | On-demand / spot | Yes | Mostly serverless | On-demand / spot |
| Billing | Usage-based (tokens / seconds / hours) | Usage-based | Usage-based | Per second or hour |
Who Is It For
- Developers who want to switch between several LLMs to compare output and cost
- Teams looking to reduce inference cost while staying on an OpenAI-compatible interface
- Product builders who want image, video, and speech generation sourced from the same API as text
- Developers who need a safe execution environment for code produced by AI agents
- Anyone who wants to run their own fine-tuned model on a GPU
Summary
Novita bundles a large model catalog behind an OpenAI-compatible API and puts GPU instances, serverless GPUs, and an agent sandbox underneath it. That layering suits early-stage products with unpredictable load: start on the model API, then shift to dedicated GPUs when the traffic pattern becomes clear. Because pricing is broken down finely by feature, it is worth running a small workload of your own before committing, and estimating the monthly figure from real consumption.