What is a Group? User Group vs Token Group Explained

A deep dive into API易’s group concept. From a user’s view it feels like ‘my group’, but what actually takes effect for every call is the group selected on the token. Includes ClaudeCode, Sora2Official, Wan dedicated-group cases and a real ticket analyzing the ‘upstream group is saturated’ 429 error.

In one sentence

A group is a “call channel” you select on a token. It decides the available models, billing multiplier, and upstream routing. From a user’s view it feels like “my own group”, but for every single call, what actually takes effect is the group selected on the token.

User view vs platform view

User view

A group is the channel I choose when creating or editing a token. It decides which models this token can call, what multiplier applies, and which upstream line it takes.

Platform view

Groups are a tool for resource management and feature surfacing — bundling similar models, dedicated capacity, and targeted discounts into one channel so billing stays precise and pricing can be differentiated.

”User group” ≠ “Token group” — don’t confuse them

A common first reaction is: “Is there a group on my account I need to switch somewhere?”

• The account level does have a “user group” concept, which decides base permission scope (whether SVIP models show up, whether enterprise fallback groups are unlocked, etc.)
• But every API call’s routing, multiplier, and model availability are decided by the group selected on the token.

When troubleshooting, first inspect the token’s “Select group” and “Fallback group” settings — don’t go hunting for “my account’s group”. See Tokens & Groups.

Case 1: Why does the ClaudeCode group exist?

Purpose: Bundle models that support the Anthropic-native /v1/messages call format into one channel, so you can use domestic coding models inside Claude Code, Cherry Studio, and other Anthropic-native clients just like calling Claude — no code changes required.

Models included:

• Full Claude lineup (official transit / AWS Claude)
• Domestic models compatible with /v1/messages, e.g. qwen3.x-max, glm-5.x, deepseek-v4

Discount:

• Default 5% off (95 折) — no action needed
• Stacks with recharge bonus (10%–20%), so real cost lands around 20% below official direct purchase

How to use:

1. Open https://api.apiyi.com/token and create or edit a token
2. Set “Select group” to ClaudeCode
3. Call from your client in Anthropic-native format

Case 2: Why do video models need a dedicated group?

Video models use billing rules (per second, per image, per duration) that are completely different from text models, and their upstream channels are independent. Groups let the special billing rules take effect precisely:

Model	Group required
Sora 2 official video	Sora2Official (per-second billing)
Alibaba Wan video series	Wan
Seedance 2 video	dedicated group (see console for the exact name)

Wrong group usually means: model unavailable (404), wrong billing, or the call is rejected outright. Make sure the token’s “Select group” or “Fallback group” includes the group that matches your target model.

Case 3: Is “Current group’s upstream is saturated” a throttle on me?

This is a high-frequency question in SaaS multi-user scenarios. It comes from a real support ticket.

Scenario:

• Developer: My tool is SaaS-style, with many users calling concurrently. Once traffic ramps, I get:

  error 429 (content-type-not-allowed): Current group’s upstream is saturated, please try again later

• I assumed the platform was throttling my concurrency — should I “set up a group” somewhere to bypass it?

The truth:

• This error is not an account-level concurrency throttle
• It says: the upstream channel mapped to that model in that group is currently busy
• Common trigger: using a model still in preview at the vendor’s side (versions named like *-preview-*), whose official capacity itself fluctuates

The right response:

1. Loosen client timeouts and retries

   Raise timeouts (e.g. 60–120s) and switch immediate retries to exponential backoff. Don’t pile on concurrent retries the moment an error fires.

2. Attach a fallback group for hot models

   On the token, add 1–2 fallback groups corresponding to the target model. When the primary is congested, traffic shifts to a backup channel and success rate improves.

3. Evaluate models for high-concurrency workloads

   If your business is latency- or stability-sensitive, neutrally evaluate more load-stable variants within the same model family on your own scenario (most vendors offer lighter, more dispersed sibling versions). Your business owns the trade-off.

We do not put a concurrency wall on customer calls. This 429 comes from the upstream channel — it is not a billing-level throttle. Retries usually recover.

How to pick a group — quick decision

Your scenario	Group to choose
Text, multimodal, NanoBanana, Veo 3.1, and most models	Default
Claude + domestic coding models in Claude Code (/v1/messages format)	ClaudeCode (5% off by default, stacks with bonus)
Sora 2 official video	Sora2Official
Wan / Seedance 2 video	their dedicated groups
Unstable high-concurrency workloads	attach 1–2 fallback groups on the token

About the “group multiplier”

The “group multiplier” shown in the console is a relative value priced in RMB, not a direct USD discount ratio — 0.14x does NOT mean “an 86% discount”. You generally don’t need to dig into it; just pick the group that matches your model. To understand multipliers and price conversion, see What is a model’s multiplier?.

Related docs

Tokens & Groups

Token roles, creation/editing, viewing code examples, group overview.

Token billing modes

Differences between pay-per-usage and pay-per-call modes.

Model multiplier

Multiplier meaning, RMB pricing unit, and USD price conversion.

Model availability

Model tiers and access by user group.