Channels & Pricing

We offer 10 independent Channels, each with transparent pricing. When you create an API Key, you bind it to one Channel, and calls are billed at that Channel's rate (the Grok Channel is billed per request — see the note below).

Channel List

Channel	Billing Rate	Best For
Claude Official (Claude Code only)	1.2x	Dedicated to the Claude Code CLI, top-quality upstream, best stability; non-CC clients get a 503 rejection
Claude Official (any client)	1.5x	Cline / Cursor calling Claude / opencode / Open WebUI / Cherry Studio / custom SDKs calling Anthropic, any client supported
Claude Official (distillation allowed) 🆕	1.5x	Same as above, and the outputs may be used for model distillation / training-data generation — pick this Channel if you need that
Claude AWS Bedrock 🆕	3.0x	AWS Bedrock upstream, high stability and high availability; all Claude models, ideal for production workflows with the strictest stability requirements
Budget Claude	0.5x	Any client, lowest price; budget tier, quality / stability relatively less reliable
Claude Sonnet Discount	0.5x ⭐	Primarily using Sonnet for coding / analysis — Sonnet series only
Codex	0.35x	OpenAI / Codex CLI / Cursor / Cline / Chatbox (supports both Responses + ChatCompletions protocols)
Codex Ultra-Budget 🆕	0.15x	All OpenAI / GPT-5.x / GPT-5.6-family models, lowest price
Gemini	0.5x	Long-form text, image understanding (Gemini 3 / 2.5, native protocol)
Grok (per-request) 🆕	per-call	xAI Grok series (grok-3 / grok-4.3 / grok-4.5 and more), billed per call not per token; /v1/chat/completions endpoint only

The Grok Channel is billed per call (not per token)

Each successful Grok call is charged a flat unit price, regardless of input/output token count:

Model	Price (per call)	Notes
grok-3	$0.014 / call	General chat (lightweight)
grok-420-fast	$0.044 / call	Stronger, faster
grok-420-fast-deepsearch	$0.044 / call	Web deep-research (slower, tens of seconds)
grok-4.3	$0.06 / call	4.3-gen general
grok-4.3-fast	$0.06 / call	4.3-gen fast
grok-4.5-latest	$0.08 / call	4.5-gen (latest)
grok-4.5	$0.18 / call	4.5-gen flagship (most capable)

• Only the OpenAI-compatible /v1/chat/completions endpoint is supported (no /v1/responses). With Cline / Cursor / Chatbox / Cherry Studio / a custom OpenAI SDK, point the base url at us and set the model to grok-3, etc.
• Because billing is per call, one call = one charge, no matter how long the conversation is. Great for short Q&A and web-search scenarios.

How the rate is calculated

Official price × Channel rate = your actual price. For example, Claude Sonnet 4.6 input at an official $3/M tokens:

• Claude Official (Claude Code only) 1.2x → $3.6/M
• Claude Official (any client) 1.5x → $4.5/M
• Claude Official (distillation allowed) 1.5x → $4.5/M
• Budget Claude 0.5x → $1.5/M
• Claude Sonnet Discount 0.5x → $1.5/M

Combined with the Top-up rate of ¥1=$1, that equals ¥1.5/M.

How to choose among the five all-model Claude Channels

The models / context length are identical — the differences are client compatibility + upstream quality + rate + whether distillation is allowed:

• Claude Official (Claude Code only) 1.2x: top-quality upstream, best stability. But the channel only accepts the official Claude Code CLI (claude-cli/X.Y.Z UA + standard metadata). Cline / Cursor / opencode / Open WebUI / custom SDKs, etc. will get a 503.
• Claude Official (any client) 1.5x: usable by any client speaking the Anthropic protocol, with good stability.
• Claude Official (distillation allowed) 1.5x 🆕: same price and same models as "any client" — the only difference is that outputs may be used for distillation / training-data generation.
• Claude AWS Bedrock 3.0x: AWS Bedrock upstream, for workflows with the most demanding stability / high availability requirements (production agents, long-running batch jobs, no tolerance for a single failure). Higher price, but the best perceived SLA.
• Budget Claude 0.5x`: any client, lowest price, but quality / stability relatively less reliable — suited for scenarios with high tolerance for occasional failures that prioritize value.

Only using the Claude Code CLI → 1.2x. Other tools needing stability → 1.5x. Using outputs for distillation / training data → distillation-allowed 1.5x. Production-critical workflows needing maximum stability → AWS Bedrock 3.0x. Chasing the lowest price and tolerating occasional failures → 0.5x budget tier.

Distillation / training-data use is allowed only on the "distillation allowed" Channel

Using model outputs to train, fine-tune, or distill your own models, or to bulk-generate training datasets, is permitted only on the Claude Official (distillation allowed) Channel.

All other Claude Channels (Claude Code only / any client / AWS Bedrock / Budget Claude / Claude Sonnet Discount) may not be used for distillation.

How to choose between the two Codex (OpenAI) Channels

The supported models are the same (GPT-5.x / all three GPT-5.6 tiers / Codex series) — the only differences are rate and route:

• Codex 0.35x: a stable, established Channel for plain-text calls (Chat Completions + Responses).
• Codex Ultra-Budget 0.15x: a new upstream Channel with the lowest price.

Want stability → Codex 0.35x. Want to save money → Codex Ultra-Budget 0.15x. For image generation, use the dedicated image channel.

Models Supported by Each Channel

Below are the exact strings you must put in the request's model field. There is no short-name fallback — writing gpt-5 instead of gpt-5.4 will get you a 503 No available accounts.

Claude all-model channels (Claude Code only / any client / distillation allowed / AWS Bedrock / Budget Claude)

claude-opus-5                    ← flagship (latest, the three "Official" channels only)
claude-opus-4-8                  ← prev-gen flagship
claude-opus-4-7
claude-sonnet-4-6                ← workhorse (recommended)
claude-haiku-4-5-20251001        ← cheap & fast
claude-opus-4-6                  ← prev gen
claude-opus-4-5-20251101         ← prev gen
claude-sonnet-4-5-20250929       ← prev gen
claude-fable-5                   ← new model (the three "Official" channels only)

# claude-opus-5 / claude-fable-5 are available only on Claude Official (Claude Code only) +
#   Claude Official (any client) + Claude Official (distillation allowed);
#   AWS Bedrock / Budget Claude do not include them yet.
# The "distillation allowed" Channel serves 8 models: opus-5 / opus-4-8 / opus-4-7 / opus-4-6 /
#   sonnet-5 / sonnet-4-6 / fable-5 / haiku-4-5-20251001 (not the older opus-4-5 / sonnet-4-5 names).
# The other models are identical across channels — the differences are client compatibility +
#   upstream quality + rate (1.2x / 1.5x / 1.5x / 3.0x / 0.5x)

Claude Sonnet Discount

claude-sonnet-4-6                ← 主力（推荐）
claude-sonnet-4-5-20250929       ← 上代

# ⚠️ 仅支持 Sonnet 系列。调 opus / haiku 会拿到 503
#   "channel pricing restriction"

Codex (0.35x)

gpt-5.6-sol                      ← 🆕 latest flagship (deepest reasoning)
gpt-5.6-terra                    ← 🆕 balanced everyday (best value)
gpt-5.6-luna                     ← 🆕 budget & fast
gpt-5.5                          ← prev gen
gpt-5.5-codex                    ← Codex coding specialist
gpt-5.4                          ← workhorse (recommended)
gpt-5.4-mini                     ← cheap
gpt-5.3-codex                    ← coding specialist
gpt-5.3-codex-spark              ← Pro-only (reasoning research preview)
gpt-5.2 / gpt-5.2-pro

# Only the models listed above work; **any unlisted model returns 503**
# Image generation (gpt-image-*) is not on Codex — use the dedicated image channel
# The Pro series (gpt-5.5-pro / gpt-5.4-pro) is Responses-API only, unavailable on the Codex path

Codex Ultra-Budget (0.15x)

gpt-5.6-sol / gpt-5.6-terra      ← 🆕 GPT-5.6 flagship/balanced (cheaper)
gpt-5.6-luna                     ← 🆕 budget & fast (cheaper)
gpt-5.5 / gpt-5.5-codex
gpt-5.4 / gpt-5.4-mini
gpt-5.3-codex / gpt-5.3-codex-spark
gpt-5.2 / gpt-5.2-pro

# Lowest price (0.15x); for image generation use the dedicated image channel

Gemini

gemini-3.1-pro-preview           ← 旗舰（带 thinking）
gemini-3-pro-preview             ← 主力
gemini-3-flash-preview           ← 快速
gemini-2.5-pro                   ← 上代旗舰
gemini-2.5-flash                 ← 上代快速

The Gemini Channel uses the native protocol

Gemini uses Google's native /v1beta/models/{model}:generateContent (or :streamGenerateContent?alt=sse) endpoint, and does not use the OpenAI / Anthropic compatible interface. Each contents item in the request must explicitly include "role": "user", otherwise Google will return 400 Request contains an invalid argument. See FAQ for details.

🎨 Image generation

Image generation is handled by the dedicated image channel, not the Codex channels. Easiest way: after logging in, use the sidebar "AI Image Generation" tool — pick a Key + prompt and generate directly, with reference-image editing + multi-platform support (OpenAI / Gemini / Antigravity).

⚡ 1M Long Context (Beta)

3 models support a 1M-token context window (default 200K):

Model	Default	1M Mode
claude-opus-4-7	200K	✅
claude-opus-4-6	200K	✅
claude-sonnet-4-6	200K	✅

How to enable it

Add a header to the request:

anthropic-beta: context-1m-2025-08-07

If you have another beta enabled at the same time (e.g. thinking), just separate them with a comma:

anthropic-beta: context-1m-2025-08-07,interleaved-thinking-2025-05-14

When calling a model that supports 1M, the header is passed through to Anthropic and 1M mode takes effect. When calling a model that does not support 1M (haiku / older opus / older sonnet), we automatically strip this header on our side and won't send an abnormal request upstream — so even if you specify the wrong model, it won't trigger Anthropic's risk controls.

Cost reminder

A single 1M-context request consumes far more tokens than 200K mode. Anthropic officially charges 2x for 1M-context input:

• Normal mode: Sonnet 4.6 input $3/M
• 1M mode (the portion beyond 200K): Sonnet 4.6 input $6/M

Converted to our pricing:

• Claude Official (Claude Code only) 1.2x → 1M mode $7.2/M
• Claude Official (any client) / (distillation allowed) 1.5x → 1M mode $9/M
• Budget Claude 0.5x → 1M mode $3/M
• Claude Sonnet Discount 0.5x → $3/M

A single call that fills the entire 1M context can cost on the order of $3–10, and multiple full-load calls within 5h will quickly consume your scheduling quota. We recommend:

• Only enable it when you truly need it (large codebase analysis, long document summarization)
• Turn it off when done
• Keep a buffer in your Top-up Balance

Which Channel Should I Choose

graph LR
    A[选择渠道] --> B{用什么工具?}
    B -->|Claude Code CLI| C{Sonnet 还是 Opus?}
    C -->|Sonnet| D[Claude Sonnet Discount]
    C -->|Opus / Haiku| E[Claude 官方<br/>仅限claude code 1.2x]
    B -->|Cline / Cursor 调 Claude / opencode / 调 Anthropic SDK| Q{要稳定还是要便宜?}
    Q -->|要稳定| F[Claude 官方<br/>不限客户端 1.5x]
    Q -->|outputs used for distillation| R[Claude Official<br/>distillation allowed 1.5x]
    Q -->|生产关键 要最稳| K[Claude AWS Bedrock 3.0x]
    Q -->|要最低价| L[Budget Claude 0.5x<br/>质量不稳定]
    B -->|Codex CLI / Cursor / OpenAI 兼容| G{要稳定还是要省钱?}
    G -->|想省钱| M[Codex Ultra-Budget 0.15x]
    G -->|要稳| N[Codex 0.35x]
    B -->|要多模态/长文档| H[Gemini]
    B -->|多渠道混用| I[创建多个 Key,<br/>各绑一个渠道]

Simplified Recommendations

• Claude Code CLI, primarily Sonnet → Claude Sonnet Discount (cheapest, Sonnet only)
• Claude Code CLI using Opus / Haiku → Claude Official (Claude Code only) 1.2x (top-quality upstream)
• Cline / Cursor calling Claude / opencode / Open WebUI over the Anthropic protocol, needing stability → Claude Official (any client) 1.5x
• Using outputs for distillation / training-data generation → Claude Official (distillation allowed) 1.5x (same price; the only Channel permitting this)
• Production-critical workflows needing maximum stability → Claude AWS Bedrock 3.0x
• Any client, chasing the lowest price, able to tolerate occasional failures → Budget Claude 0.5x
• Codex CLI / Cursor calling GPT in plain text → Codex 0.35x
• Want maximum savings → Codex Ultra-Budget 0.15x
• Not sure → Codex Ultra-Budget (best compatibility, works with almost any tool that supports the "OpenAI API" + lowest price)

Top-up Rate

¥1 = $1 (Top-up rate 1.0x). For example:

• Top up ¥100 → wallet Balance $100
• One full Claude Sonnet 4.6 conversation costs about $0.018 (including 1k input + 500 output tokens, 0.5x discount Group)
• That works out to roughly 5500 medium-length conversations per ¥100

See Top-up & Plans for details.

Service Rate Limiting

To keep the service stable over the long term, our scheduling system automatically controls the call volume per unit of time. When requests come in too densely within a short window, some requests will temporarily return 429 and automatically recover within 1–5 minutes.

If your workflow is sensitive to occasional 429s, we recommend:

• Implementing simple exponential-backoff retries in your client
• Spreading out usage (avoiding the 19:00–23:00 peak)
• Rotating across multiple Keys for important tasks

See FAQ for details.