Novita AI vs Runcrate

One endpoint for 200+ text, image, audio and video models, callable with existing OpenAI client libraries

Qwen, DeepSeek, GLM, Kimi, MiniMax, Llama, Gemma and Nemotron families served serverlessly with per-token billing

Cache-read pricing published separately from standard input pricing on most chat models

Per-second serverless GPU jobs, on-demand instances, and bare-metal clusters with H100 and H200 hardware

Dedicated deployments for workloads that need reserved capacity rather than shared serverless throughput

Isolated runtimes for agent workloads that call models and execute tools, billed per second

Token-billed API calls and GPU hours draw on one credit balance with a single invoice, auto-recharge, and credits that do not expire

Chat, image, video, text-to-speech, speech-to-text and embedding models served behind OpenAI-compatible REST endpoints, so existing clients switch by changing the base URL

Bare-metal instances with SSH, Docker and custom images, provisioned in about 60 seconds

An unauthenticated endpoint publishes the live GPU rate card, refreshed every few minutes

Reserved bare-metal from 16 to 128+ nodes with NVLink fabric and managed Slurm job submission

Python and TypeScript SDKs, a Vercel AI SDK provider, a CLI for instance and volume management, and an MCP server for controlling the platform from AI assistants