One Key for every mainstream model
All models are billed at the published official price; discounts follow the notice on this site. Check the live catalog with your Key:GET /v1/models
DeepSeek-V4-Pro (1.6T/49B active) matches top closed models, with native 1M context.
Pinned 0813 snapshot of V4 Pro — same capability and price, for stable version targeting.
Lightweight high-speed model — lowest cost with fast responses, built for high-volume calls.
A 2.4-trillion MoE open flagship that can code on its own and deliver complete projects spanning 10+ days.
0902 snapshot of 3.8 Max with long context, for pinned-version deployments.
Only 6B parameters activated per token, yet a 1M context.
Vision model with text, image and video input for image and document understanding.
The first 3-trillion-class open model — 2.8T parameters, native vision, million-level context, built for long-horizon coding and knowledge work.
All of the gains come from post-training; 50% better at coding than 5.2, the strongest open model.
The fast, affordable version of GLM-5.3, suited to high-volume and low-latency workloads.
MiniMax-M3 supports image understanding: read screenshots and documents, OCR, analyze photos. It shares the same API key and balance as the chat models.
Standard narration voice; long-text segmentation; 22.05 kHz WAV output.
Conversational voice for narration and support scenarios.
Bright, lively voice for short-form content.
Clone a voice from a short sample, zero-shot.
Chinese and English transcription for meeting notes and subtitles.
Embedding model tuned for Chinese semantic search and clustering.
English-oriented embedding model, long-text friendly, ideal for RAG retrieval.
How billing works
Usage-based, metered per request. No subscription, credit never expires.
The console itemises every call — input, output, cached and reasoning tokens are all visible. Live prices can be verified with GET /v1/models .
Cache hits cost less: as low as ¥0.30 per million (about 1/30 of the input price) — save about 97% on repeated prefixes