Core features

primerouter is built around three goals: unified access, transparent billing, provable audit. Below is what's user-visible.

Models & providers

Multi-provider, drop-in

Coverage spans both Western and Chinese frontier models, including but not limited to:

• International: OpenAI (GPT-4o, o-series), Anthropic (Claude family), Google (Gemini family), Mistral, Cohere, xAI (Grok), Meta Llama
• Chinese: DeepSeek (V3 / R1), Alibaba Qwen, Zhipu GLM, Moonshot Kimi, MiniMax, ByteDance Doubao, Baidu Ernie, Tencent Hunyuan, iFlytek Spark
• Specialty: Ollama (local), SiliconFlow, Dify, Coze, Replicate

Full model list and pricing in Pricing → Models.

Native multi-format

Not just OpenAI-compatible — native formats of Claude and Gemini are also first-class, meaning the Claude SDK / Google AI SDK can connect directly:

Format	Endpoint	Use
OpenAI Chat Completions	/v1/chat/completions	Most widespread
OpenAI Responses	/v1/responses	New OpenAI format
OpenAI Embeddings	/v1/embeddings	Vector embeddings
OpenAI Images / Audio	/v1/images/..., /v1/audio/...	Multimodal
Anthropic Messages	/v1/messages	Native Claude
Google Gemini	Native Gemini paths	Native Gemini

Format conversion happens in the middle layer. You can swap upstream without code change: call Claude in OpenAI shape, call Gemini in Claude shape, etc.

Multimodal

Modality	Capability
Text chat	Chat completions / streaming / multi-turn / tool use / reasoning
Embeddings	OpenAI / Cohere Embed
Image generation	DALL·E, Midjourney, Jimeng
Audio	TTS / STT (OpenAI, Volcengine, MiniMax)
Video generation	Sora, Kling, Hailuo, Vidu, Jimeng, Doubao (async tasks)
Music generation	Suno
Reranking	Cohere, Jina, Alibaba

Long-running modalities (video, music) go through an async task system: submit → pre-charge → poll → settle. See Features → async tasks.

Routing & availability

Smart routing

• Weighted random: distribute by per-channel weight
• Priority retry: high priority first; on failure, fall back to next tier
• Cross-group retry: auto tokens can iterate across user groups
• Channel affinity: same user × model is sticky to one channel by default

Multi-key fault isolation

Each upstream channel can hold many keys (round-robin, random, or weighted). A single key failure is auto-skipped, leaving the others unaffected.

Self-healing

• AutoBan: a channel that keeps failing is disabled, then auto-recovered
• Realtime latency monitoring
• Balance alerts on upstream accounts
• Periodic health checks

Billing & payments

Token-based, no subscription

• Real tokens × model price = your bill. No monthly fee, no bundle, no minimum.
• The usage field is byte-identical to upstream. Each request can be drilled down to the upstream ledger.
• Pre-charge on call, settle on completion, refund on failure — no double charge, no missed accounting.

Multiple top-up methods

• Crypto (differentiator): USDT / USDC etc. across multiple chains (EVM, Aptos). On-chain confirmation triggers credit. See Crypto Payment.
• Cards / Apple Pay / Google Pay (Stripe)
• Local payment methods (Waffo, 162 countries, 480+ methods)
• Redemption codes / referral rewards

PrimeRouter Chat

More than an API gateway — PrimeRouter Chat is a chat product for people who don't write code, on its own domain https://chat.primerouter.ai:

• Sign up and sign in with email; the same account and balance as the console, pay as you go
• Pick the model and how hard it thinks; web search with cited sources
• Streaming replies you can stop or regenerate; the server keeps generating if you switch away or lose the connection, and you pick up where it is
• Send images (some models also read PDFs), tap an image to view it full screen
• Edit your last message and regenerate; tap any reply to see what it cost
• Install it to your phone's home screen and read past conversations offline; on a computer it gets a sidebar layout

The console's top-bar 「聊天」 (Chat) entry, once an admin turns it on, opens it signed in.

Playground

Visual API debugging: model picker, parameter sliders, request body editing, real-time SSE viewer, one-click code-snippet copy, save test configs.

Console

Available to every logged-in user: dashboard, personal, tokens, top-up, usage logs, Playground, models, subscription, redemption, Midjourney, async tasks; the top-bar Chat entry opens PrimeRouter Chat. See Features.

Admin

Administrators have access to upstream channel management, model deployment, user management, group pricing, and system settings. These are operational concerns and are intentionally not covered in this user documentation.

Security

• OAuth: GitHub, Discord, OIDC, LinuxDO, etc.
• Passkey (WebAuthn) + 2FA TOTP
• CAPTCHA (Cloudflare Turnstile / hCaptcha)
• Per-key model allow-list, IP allow-list, quota cap, expiry
• SSRF protection, two-tier rate limiting, CORS, request trace ID

Internationalization

Docs and product UI both support 7 languages: 简体中文, English, 繁體中文, 日本語, Français, Русский, Tiếng Việt.

Next

• Architecture
• 5-minute start