APIs for LLMs, image, video, audio generation, and GPU cloud
Novita AI is an API platform for LLM inference, image/video/audio generation, GPU cloud, and agent sandboxes.
AI Panel Score
6 AI reviews
Reviewed
Facts checked against novita.ai
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.Developers interact with Novita AI primarily through its REST API, authenticated with a bearer token, using either the platform's native API base URL or an OpenAI-compatible endpoint for LLM tasks like chat completions, text completions, embeddings, and reranking. Beyond text, the API reference covers image generation and editing (including Qwen Image txt2img and edit), video generation (Kling v3.0 text-to-video and image-to-video variants, Minimax Hailuo 2.3), audio generation (Minimax Speech, Minimax Music, Fish Audio text-to-speech and voice cloning), and third-party integrations such as Exa for search and answer retrieval, and HeyGen for video translation.
The platform also offers GPU cloud infrastructure with a separate API surface for provisioning and managing compute instances: creating, starting, stopping, restarting, migrating, and upgrading GPU or CPU instances, managing VPCs, network storage, container registry authentication, image prewarming, and templates. A newer "GPUS V2" API set adds endpoint management, job tracking, and autoscaling-related controls like auto-migrate and autorenew settings. Account-level features include usage-based billing queries, budget controls that can restrict spending per API key or team member, and IP/model access policies scoped to individual keys.
Generates vector embeddings from text via a dedicated API endpoint.
Provides an OpenAI-compatible endpoint for creating chat completions and text completions with large language models.
Reranks a set of documents or passages against a query for improved retrieval relevance.
Retrieves instance metrics and job status information for GPU compute workloads.
Provides endpoints to query account balance, transactions, monthly bills, usage-based billing, and fixed-term billing details.
Checks and updates budget configurations for API keys and members, including verifying whether a key or member has sufficient budget for a given cost.
Allows creating, listing, retrieving, and canceling batch jobs for processing large volumes of LLM requests asynchronously, including file upload and management.
Enables creating, editing, starting, stopping, restarting, migrating, and deleting GPU (and CPU) compute instances for heavy workloads.
Provides APIs to create, edit, list, and delete network storage volumes and virtual private clouds (VPCs) for GPU instances.
Supports creating and managing instance templates, saving custom images, and prewarming images for faster deployment.
Manages authentication credentials for pulling container images from private registries into GPU instances.
Exposes dedicated APIs for third-party models covering video generation (Kling, Minimax Hailuo), image generation/editing (Qwen Image), text-to-speech and voice cloning (Fish Audio, Minimax Speech), music generation (Minimax Music), web search (Exa), and video translation (HeyGen).
Configures per-key IP access policies and model access policies to control which models and networks an API key can use.
Usage-based access to hosted LLM, image, audio, video, and AI search models via API, billed per million tokens (input/output) with cache-read discounts on select models.
Dedicated compute resources for hosting models with predictable, reserved capacity rather than shared serverless infrastructure.
Sandbox environment for building and running AI agents, priced separately from model API usage.
On-demand and reserved GPU resources for training, fine-tuning, and inference workloads, priced by GPU type and usage.
One API for 200+ models plus GPU cloud, priced by the token instead of by feature lock-in.
“Novita bundles LLM inference, image/video/audio generation, and raw GPU instances behind one bearer token. The pricing page is unusually specific for this category, which is a good sign.”
Two hundred models, one OpenAI-compatible endpoint, and real per-token numbers on the pricing page: $0.14/Mt input on Gemma 4 31B, $1.74/Mt on DeepSeek V4 Pro. That specificity matters. Most GPU cloud vendors hide behind 'contact sales' for everything; Novita does that for GPU and dedicated tiers but not for its core model API.
The range is the pitch. Kling and Minimax for video, Fish Audio for voice cloning, Exa for search, plus bare metal and serverless GPU for the heavier workloads. That's a lot of surface area for one vendor to keep current across third-party model updates.
Tradeoff: breadth over depth. You get one bill and one key instead of five vendor relationships, but you're also betting Novita keeps pace with every upstream model it wraps. No free trial to de-risk that bet before you commit spend.
Bare metal, serverless GPU, and 200+ third-party model APIs in one billing surface is broader than most single-purpose inference APIs offer.
OpenAI-compatible endpoints ease portability, but per-key IP and model access policies are the extent of security detail I could find.
Pay-per-token pricing with no commitment and free-tier models like Ling 3.0 Flash means testing starts immediately.
Consolidating model APIs and GPU compute under one key advances multi-model product work, not just cost-cutting on a single vendor.
Detailed API reference, changelog, and a GPUS V2 surface with autoscaling controls point to active build-out.
Teams building multi-modal products who want one API and one bill instead of five vendor contracts.
Skip it if you need a no-commitment trial before touching production spend.
One bearer token, 200+ models, and a GPU fleet underneath — the aggregation bet is real, the lock-in is subtler.
“Novita AI stitches an OpenAI-compatible LLM endpoint to third-party model APIs and raw GPU provisioning under one billing surface. That's a genuinely useful architecture for a small team, but it puts a broker layer between you and every upstream model.”
The API surface is wide: chat completions, embeddings, and rerank on an OpenAI-compatible path, plus dedicated endpoints for Kling, Minimax, Qwen Image, and Fish Audio, plus GPU instance management with VPCs, network storage, and container registry auth. That's not a wrapper — it's an actual control plane spanning inference and infrastructure. Per-token pricing is granular (Gemma at $0.14/Mt input, DeepSeek V4 Pro at $1.74/Mt input) which signals real cost engineering, not a markup slapped on OpenAI's meter.
Domain fit is strong for teams that want one bearer token instead of five vendor SDKs. The risk: budget controls and access policies live at Novita's layer, so if pricing or model availability shifts upstream, your abstraction is now your constraint, not your convenience.
Three years out, this is a fine bet for prototyping and mid-scale production, weaker for anyone who needs contractual guarantees direct from a model vendor. Bare Metal and dedicated endpoints give an escape hatch from shared serverless if you outgrow it. Integration surface is genuinely broad; portability is the tradeoff you're signing up for.
Sits between a pure inference API and a GPU cloud, competing on breadth (Kling, Minimax, Fish Audio, bare metal) rather than depth in any one lane.
OpenAI-compatible endpoint plus 200+ third-party model APIs matches how teams actually build multi-model pipelines today.
Bearer-token auth, VPCs, container registry auth, and network storage cover most of a GPU deployment workflow in one API set.
Convenient now, but budget controls, access policies, and billing all sit inside Novita's layer rather than the upstream vendors.
GPUS V2 autoscaling, batch inference, and templates go past a basic model proxy into real infra tooling.
Engineering teams that want one API and one bill across LLMs, media generation, and GPU compute instead of stitching together five vendors.
Avoid if you need direct contractual terms or SLAs with a specific model vendor rather than a broker sitting in between.
Per-token pricing is public. GPU and dedicated tiers still need a sales call.
“DeepSeek V4 Pro at $1.74/Mt input, $3.48/Mt output. Gemma 4 31B at $0.14/$0.4. Real numbers on the model side, silence on GPU and sandbox pricing.”
200+ models, one API. Token pricing published per model — DeepSeek V4 Pro $1.74/Mt in, $3.48/Mt out. That's usable procurement math without a call.
GPU Instances, Dedicated Endpoints, Agent Sandbox: all four listed "contact sales." No commitment on serverless, which helps TCO — pay per token, scale to zero on GPU jobs. But mixed workloads (LLM + GPU training + agent sandboxes) mean stitching two pricing models, one transparent, one opaque. Budget controls per API key are a real cost-governance feature — most usage-based vendors skip this.
Billing looks clean: usage queries, monthly bills, fixed-term billing all exposed via API. No free trial, no free plan, so first invoice is real spend, day one. Batch inference gets a 50% introductory discount — check if that's permanent before modeling year 3.
Bearer-token auth, budget controls per key, and billing query endpoints suggest low procurement friction for the API side.
Serverless is pay-as-you-go with no lock-in; dedicated/GPU terms aren't published so renewal terms are unknown to me.
Per-token model pricing is public with real numbers; GPU, dedicated, and sandbox tiers all say contact sales.
Per-model token cost plus usage/billing query APIs make cost-per-request measurable, which is more than most model marketplaces offer.
No-commitment token billing keeps entry cost low, but GPU/dedicated costs are unknowable without a quote.
Teams that want transparent per-token model pricing and are comfortable negotiating separately for GPU or dedicated capacity.
You need a single, fully public price sheet across model APIs and GPU infrastructure before talking to anyone.
OpenAI-compatible endpoint plus GPU provisioning in one API surface — rare combo, uneven polish
“Novita bundles 200+ model APIs with actual GPU instance management under one bearer token. The breadth is real, but pricing opacity across GPU tiers means budgeting requires a sales call before you write a line of code.”
OpenAI-compatible chat completions, embeddings, and rerank in one endpoint means dropping this in as a base_url swap for anything already built on the OpenAI SDK. That's the right call — zero migration tax for the common path. Token pricing is transparent too: DeepSeek V4 Pro at $1.74/Mt input, Gemma 4 31B at $0.14/Mt input, published per-model, no sales call needed for inference.
GPU instances are a different story. Creating, migrating, autoscaling via 'GPUS V2' — real infra primitives, VPCs and network storage included — but every GPU tier (Serverless, Dedicated, Bare Metal) says contact sales. That's a context switch mid-project: prototype on the token API in five minutes, then stall on GPU pricing for a training job.
Batch API with file upload for async LLM jobs is a nice touch for eval pipelines. Per-key IP and model access policies suggest someone building this has run multi-tenant workloads before. Docs reference for third-party models (Kling, Fish Audio, Exa) reads like it's tracking upstream API changes closely, not just wrapping them once and forgetting.
Token-priced LLM calls work immediately; GPU provisioning likely stalls on a sales conversation before first deploy.
API reference covers granular endpoints (batch jobs, VPC, container registry auth) suggesting docs written against real usage, not marketing copy.
Two pricing models under one platform — transparent per-token for inference, opaque contact-sales for GPU/Dedicated/Bare Metal — adds planning overhead.
Budget controls per API key, IP/model access policies, and GPUS V2 autoscaling give room to grow past hello-world calls.
OpenAI-compatible endpoint means existing SDK code points here with a base_url change, no rewrite.
Teams already on an OpenAI-compatible stack who also need occasional GPU bursts for fine-tuning or video/audio model workloads.
Avoid if you need transparent, self-serve GPU pricing without a sales call before you can estimate monthly cost.
200+ models behind one API key, but this is a tool for people who read docs first
“Novita AI bundles LLMs, image, video, audio, and GPU cloud into one bearer-token API with an OpenAI-compatible layer. It's built for developers wiring things into pipelines, not for anyone expecting a friendly dashboard first impression.”
This isn't a product you click around in. It's a product you curl. There's an OpenAI-compatible endpoint for chat, embeddings, and reranking, which means if you've built against OpenAI before, day one is copy-paste-and-swap-the-base-URL easy. That's a real kindness to developers, and it's the best onboarding shortcut a platform like this can offer.
Past that first swap, it gets homework-y fast. You're managing GPU instances, VPCs, network storage, container registry auth, and a separate 'GPUS V2' surface with autoscaling knobs, on top of picking between DeepSeek at $1.74/Mt input or Gemma at $0.14/Mt. That's a lot of surface area for month one, though by month three, if you're the kind of person who lives in API references, it probably feels like power rather than clutter.
No dashboard screenshots to judge daily polish by, no free trial to test the water first — you're contact-sales or nothing on every tier. Budget controls per API key are a nice touch for teams who've been burned by runaway usage bills before. Mobile isn't really the point here; this lives in your backend, not your pocket.
Budget controls and per-key access policies show real thought about how teams actually manage spend day to day.
Chat completions are trivial to start, but VPCs, bare metal, serverless GPU, and 200+ models add real depth to climb through.
This is an API and GPU cloud platform; there's no indication of a mobile app, which is a category norm rather than a specific failing.
The OpenAI-compatible endpoint means anyone with existing chat-completions code can swap in fast, though GPU and sandbox setup is a heavier lift.
GPU metrics and job monitoring endpoints exist, but I couldn't find public uptime or status-page data to judge track record.
Developers who want one API key covering LLM, image, video, audio, and GPU workloads without juggling five vendors.
You want to see a dashboard or try before you buy, since there's no free trial and pricing is contact-sales.
200+ models, one API key, and a title tag that literally says 'KimiKimiKimi'.
“Novita bundles LLM inference, GPU cloud, and third-party media models behind one OpenAI-compatible endpoint. Breadth is real. Whether any of it is defensible is another question.”
DeepSeek V4 Pro at $1.74/Mt input, Gemma 4 31B at $0.14/Mt — specific numbers, which I like. Beats vague 'enterprise pricing' pages. The OpenAI-compatible endpoint means switching in is cheap. Switching out is the real question: chat completions port easily, but GPU templates, VPC configs, and budget controls tied to Novita's API keys don't move with you.
The model list — Kling, Minimax, Qwen Image, Fish Audio, Exa, HeyGen — reads like an aggregator play. That's not a knock. It's the pitch: one bill instead of six vendor relationships. But aggregators live and die on margin and uptime guarantees I couldn't find published anywhere.
GPUs V2 with autoscaling and job tracking suggests active build-out, not a stale shell. Fine signal. The vendor site's own title tag reading 'KimiKimiKimi' is the kind of sloppy detail that makes we wonder what else is unpolished.
Bundling LLM inference, GPU cloud, and media-gen APIs under one key is a real convenience play, not obviously unique against single-purpose model APIs.
OpenAI-compatible chat endpoint eases entry and partial exit, but GPU templates, VPCs, and budget policies are Novita-specific lock-in.
A GPUS V2 API set with autoscaling and job tracking points to active development beyond a static launch.
'AI-Native Cloud for Builders and Agents' is aspirational, but the 200+ model count and per-token pricing examples are concrete and checkable.
Named APIs (embeddings, rerank, batch processing) and specific model pricing back up the claims better than most usage-based platforms manage.
Teams that want one API key and one bill across LLMs, GPU compute, and media generation models.
Skip it if you need transparent published pricing on GPU or dedicated tiers before you'll even talk to sales.
Common questions answered by our AI research team
Model API pricing is usage-based, billed by the token rather than by the hour. For example, DeepSeek V4 Pro costs $1.74/Mt input and $3.48/Mt output, while Gemma 4 31B costs $0.14/Mt input and $0.4/Mt output, with pricing varying by model and context length.
Novita AI supports 200+ models across LLM, image, audio, video, and vision categories, all served through a single serverless API.
Yes. Novita AI offers full-control GPU Instances that are dedicated and yours in seconds, plus Serverless GPU for job-based scaling and Bare Metal for dedicated physical GPU clusters.
Yes, Novita AI exposes an OpenAI-compatible endpoint for chat completions, embeddings, and reranking, alongside dedicated model APIs for third-party models like Kling, Minimax, Qwen Image, and Fish Audio.
Serverless GPU lets you submit a job with no instances to provision — Novita allocates GPU resources automatically, scales to zero when idle, and bills per execution. Bare Metal gives dedicated physical GPU clusters with zero abstraction overhead for large-scale inference, training, and enterprise workloads needing full hardware control.
Novita AI is a cloud infrastructure provider offering model APIs, GPU instances, and serverless GPU deployment for AI workloads.
