> ## Documentation Index
> Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Groups: Official Relay by Default, Reverse as a Supplement

> APIYI is official-relay by default: an enterprise-grade AI API provider focused on quality and availability. This guide covers Default, SVIP, ClaudeCode, Claude_Enterprise, Claude_Reverse, Codex_Reverse, Gemini_Reverse, image2Enterprise, NB-Enterprise, and more, with multipliers, discounts, and model coverage.

**APIYI is official-relay by default.** As an enterprise-grade AI API provider, our default channels are vendor official APIs and official cloud marketplaces, focused on quality, availability, and long-term stability. Reverse groups are a supplement for personal learning, cost-sensitive workloads, and non-critical tasks.

A group is the **call channel attached to a token**. It decides which models the token can use, which multiplier applies, and which upstream route the request takes. You do not need to memorize the table; follow three steps:

**1. Check which groups support the model**

Use Model Pricing in the console or the token editor to confirm the model's groups. If the model is not in the selected group, the request returns "no available channel for this model".

**2. Match the group to the workload**

Production, enterprise, and coding Agents should prefer official relay or enterprise groups. Personal learning, simple tasks, and cost-sensitive workloads can use reverse groups.

**3. Compare price and set a fallback**

Compare the group multiplier and the effective discount after recharge bonuses. If the primary group is unstable, configure a fallback group on the token.

> [!NOTE]
>
> **A multiplier is a billing coefficient, not the final discount.** For text models, the actual unit price is model multiplier × group multiplier. Image and video models are billed per call or per second; the console's Model Pricing is authoritative.

## How to use a group: select it when creating a token

A group is not a separate switch; it is a **field you select when creating or editing a token**. Changing groups does not require code changes: change the token's group, and the model name and endpoint stay the same.

**1. Open the tokens page**

Go to the [Tokens page](https://api.apiyi.com/token) in the console and create a token or edit an existing one.

**2. Select a group**

Choose the target group under "Select group". If you are unsure, start with `Default`.

**3. Add fallback groups**

For higher success rates, add 1–2 backup groups under "Fallback group"; traffic falls back automatically when the primary group fails.

**4. Save and call**

Save the token and send requests as usual. The group affects routing and billing only; the model name and endpoint do not change.

> [!TIP]
>
> If you need multiple primary groups, use separate tokens by workload so logs and costs are easier to review. For the full procedure, see [Tokens & Groups](https://docs.apiyi.com/en/faq/token-and-groups.md).

## Official relay and reverse: official by default, reverse as a supplement

**Official relay** means resources forwarded through a vendor's official API or official cloud marketplace, such as the Anthropic API, Claude on AWS, Google AI Studio / Vertex, Volcano Engine, and Alibaba Cloud. **Reverse-engineered** means access obtained through unofficial interfaces such as the web app, CLI, or IDE, with APIYI pooling and scheduling the resources.

APIYI is official-relay by default. The official-relay groups come first; the reverse section explains where reverse fits.

### Official relay: the default channel

| Group | Source | Model coverage | Positioning |
| - | - | - | - |
| `Default` | Mostly vendor official APIs / official cloud marketplaces | Most text, multimodal, image, and video models on the platform | Default choice; official price for text |
| `SVIP` | Mostly vendor official APIs / official cloud marketplaces | Everything in `Default` plus scarce Pro models | For `gpt-5.2-pro` / `gpt-5.5-pro` / `o3-pro` and similar |
| `ClaudeCode` | Anthropic official API + AWS Claude | All Claude models plus coding models compatible with `/v1/messages` | Claude Code / Anthropic native, 5% off |
| `Claude_Enterprise` | Anthropic official direct, excluding AWS Claude | All Claude models | Enterprise production / emergency, 1.2× |
| `CodexResponses` | Official APIs, exposed through a Responses-compatible channel | Models supporting the Responses API: GPT / Grok / DeepSeek / Seed / GLM | Codex / Responses API |
| `image2Enterprise` | OpenAI official image channel | GPT Image 2 / 2.5 / 2.6 flare and sunburst | GPT Image enterprise fallback, 1.2× |
| `NB-Enterprise` | Enterprise high-availability fallback | Nano Banana / Pro / 2 / 2.1 and more | High-quality Nano Banana fallback, `Default` × 1.4, main models still about 50% of official |
| `SeeDance2` | Volcano Engine official resources | All Seedance 2.0 / 2.5 models | Seedance video, supply + face whitelisting/asset library |
| `Wan&HappyHorse` | Alibaba Cloud official resources | Wan 2.6 / 2.7 / 3.0, HappyHorse 1.0 / 1.1 | Alibaba video, dedicated pricing |

> [!NOTE]
>
> `Default` / `SVIP` are **mostly official relay but also contain some reverse-engineered image models** such as `gpt-image-2-all`, `gpt-image-2.6-all`, and the `-vip` series. For image models, check the exact model and group rather than the group name alone.

### Reverse: cost-effective supplement

If you are doing personal learning, cost-sensitive work, or non-critical tasks, you can use the three reverse groups below. Production, enterprise, and complex Agent workloads should still prefer official relay.

| Group | Source | Coverage | Best for |
| - | - | - | - |
| `Claude_Reverse` | Reverse-engineered Claude resources | 24 Claude models, excluding newer releases such as `claude-fable-5-2` and `claude-haiku-5-5` | Personal learning, non-production |
| `Codex_Reverse` | Reverse-engineered Codex resources | 9 models including `gpt-5.6-*`, `gpt-6-*`, and `gpt-6.1-*` | Cost-sensitive Codex, non-production |
| `Gemini_Reverse` | Reverse-engineered Gemini CLI / Antigravity resources | 11 models from Gemini 2.5 to 3.8, excluding Gemini 4 | Gemini CLI, learning, non-production |

> [!NOTE]
>
> Apart from Anthropic, OpenAI, and Google, vendors do not have a separate text reverse group; their text models use the official-relay `Default` / `SVIP` groups. Reverse groups are a supplement and do not promise the same SLA, concurrency, or feature completeness as official relay.

> [card] [What Is the Difference Between Official-Relay and Reverse-Engineered APIs?](https://docs.apiyi.com/en/faq/official-vs-reverse-api.md)
>
> A dimension-by-dimension comparison of stability, protocol compatibility, feature completeness, caching, concurrency, and long-term sustainability, plus five acceptance metrics.

## All groups at a glance

The table is ordered by general, Claude, OpenAI, Gemini, image, and video groups. Multipliers and discounts were verified on 2026-11-14 (UTC+8); groups and prices can change with upstream adjustments, so the console is authoritative.

| Group | Type | Model coverage | Group multiplier | Effective price vs official | With 10% bonus | With 20% bonus |
| - | - | - | - | - | - | - |
| `Default` | General | Most models on the platform; includes reverse image models such as `-all` / `-vip` | 1 | Official price (text) | \~91% | \~83% |
| `SVIP` | General | Everything in `Default` plus scarce Pro models | 1 | Official price (text) | \~91% | \~83% |
| `ClaudeCode` | Official relay | All Claude models plus coding models compatible with `/v1/messages` | 0.95 | 95% | \~86% | \~79% |
| `Claude_Enterprise` | Official relay | All Claude models; pure official relay, no AWS Claude | 1.2 | 1.2× official | \~109% | Official price |
| `Claude_Reverse` | Reverse | 24 Claude models, excluding the newest releases | 0.5 | 50% | \~45% | \~42% |
| `CodexResponses` | Official relay | Models that support the Responses API: GPT / Grok / DeepSeek / Seed / GLM | 1 | Official price | \~91% | \~83% |
| `Codex_Reverse` | Reverse | 9 models including `gpt-5.6-*` / `gpt-6-*` / `gpt-6.1-*` | 0.5 | 50% | \~45% | \~42% |
| `Gemini_Reverse` | Reverse | 11 models from Gemini 2.5 to 3.8, excluding Gemini 4 | 0.5 | 50% | \~45% | \~42% |
| `image2Enterprise` | Official relay | GPT Image 2 / 2.5 / 2.6 flare and sunburst | 1.2 | 1.2× official | \~109% | Official price |
| `image2_OSS` | Output format | GPT Image 2/2.5/2.6 `-all` / `-vip` / flare / sunburst | 1 | No markup (by model) | By model | By model |
| `NB-Enterprise` | Enterprise fallback | Nano Banana / Pro / 2 / 2.1, `gemini-3-pro-image`, and more | 1.4 (vs `Default`) | \~50% of official (main models) | \~45–50% | \~42–47% |
| `NB_OSS` | Output format | Nano Banana Pro / 2 / 2.1, `gemini-3-pro-image`, and more | 1 (same as `Default`) | from \~37% of official | from \~34% | from \~31% |
| `SeeDance2` | Video | All Seedance 2.0 / 2.5 models, including fast and mini | 0.18 | \~1.26× official | \~1.15× | \~1.05× |
| `Wan&HappyHorse` | Video | Wan 2.6 / 2.7 / 3.0 and HappyHorse 1.0 / 1.1 | 0.14 | \~98% | \~89% | \~82% |

> [!NOTE]
>
> **How to read the percentages**: they are the share of the official price, not a percentage off. Below 100% means cheaper than the official price; above 100% means more expensive. "With 10% / 20% bonus" is the effective price after a recharge bonus, calculated as the group multiplier ÷ 1.1 or ÷ 1.2. The discount columns for `Default` / `SVIP` refer to text models; image models (including Nano Banana and GPT Image) are priced by the specific model, as explained below.
>
> **The 0.18 / 0.14 video multipliers are not direct discounts.** They are RMB pricing conversion coefficients. `SeeDance2` is nominally about 1.26× the official price and about 1.05× after a 20% bonus. `Wan&HappyHorse` is about 98% of the official price by default and about 82% after a 20% bonus. Recharge bonuses follow the current promotion rules.

> [!NOTE]
>
> **`NB-Enterprise` is not 1.4× the official price. It is the high-quality enterprise fallback for Nano Banana.** Its 1.4 multiplier is relative to `Default`: Nano Banana on `Default` is already about 37–50% of the official price depending on the model, so `NB-Enterprise` is about 50% of official for the main models and about 42–47% after a 20% bonus — still below half price. At the lowest `Default` tier (\~33% after bonus), the enterprise group is about 46%; before bonus, the main models are about 50–56% of official. `NB_OSS` is priced the same as `Default`. The console is authoritative.

> [!NOTE]
>
> **`SeeDance2` competes on supply and added capabilities, not price.** The default price is about 1.26× official; after a 20% bonus for large recharges it is about 1.05×, still a \~5% premium. From a tax perspective (issuing a 6% VAT special invoice), this channel does not make money on supply — it exists for customer convenience. The real advantages are **face whitelisting (virtual faces + real-person authorization)** and the **free asset library**: official Seedance blocks real-person images by default, and whitelisting enables them (real-person reference assets still go through the asset library); the asset library required for character consistency is a separately purchased add-on on the official side (starting in the tens of thousands of RMB per year), while APIYI includes it free with the Seedance API. See [Seedance 2.0 / 2.5 overview](https://docs.apiyi.com/en/api-capabilities/seedance2/overview.md) for the full breakdown.

> [!TIP]
>
> For scarce models such as `gpt-5.2-pro`, `gpt-5.5-pro`, and `o3-pro`, the token must use `SVIP`; these models are not in `Default`.

## Transparent pricing and recharge bonuses

APIYI has **no separate business-negotiation policy**: prices are the same for every customer. **The only discount is the recharge bonus program.** We recommend that enterprise customers participate as well; it is more cost-effective over the long run. Recharge more, get more — reliable long-term service that helps your business succeed.

> [card] [Do enterprise customers get separate pricing?](https://docs.apiyi.com/en/faq/enterprise-vs-individual.md#do-enterprise-customers-get-separate-pricing)
>
> No. Pricing is public and identical for all customers, which removes the back-and-forth of price negotiations.

> [card] [Recharge bonus program](https://docs.apiyi.com/en/faq/recharge-promotions.md)
>
> First-recharge bonus and tiered bonuses of 10%–20%. The more you recharge, the higher the bonus — better value over time.

## Choosing by model vendor

Start with whether the vendor has a separate reverse group, then check whether you need a specific protocol or output format, and finally decide between production stability and cost:

1. **Does the vendor have a separate reverse group?** Only Anthropic, OpenAI, and Google do. For every other vendor, text models use the official-relay `Default` / `SVIP` groups; image and video models use their dedicated groups.
2. **Do you need a specific protocol or output?** Claude native → `ClaudeCode`; Responses → `CodexResponses`; image URL → `image2_OSS` / `NB_OSS`; video → `SeeDance2` / `Wan&HappyHorse`.
3. **Production or cost first?** Production and enterprise workloads use official relay or enterprise groups. Personal learning and non-critical workloads can use reverse groups.

| Vendor | Production default | Dedicated options | Cost / learning | Notes |
| - | - | - | - | - |
| Anthropic | `Claude_Enterprise` / `Default` | `ClaudeCode` (0.95) | `Claude_Reverse` (0.5) | `Claude_Enterprise` is pure official relay without AWS Claude; `ClaudeCode` includes AWS Claude |
| OpenAI | `Default` / `SVIP` | `CodexResponses`, `image2Enterprise`, `image2_OSS` | `Codex_Reverse` (0.5) | Image `-all` / `-vip` are reverse models served by `Default`; some URL models use `image2_OSS` |
| Google | `Default` | `NB-Enterprise`, `NB_OSS` | `Gemini_Reverse` (0.5) | Gemini 4 is only in `Default`; `Gemini_Reverse` goes up to 3.8 |
| ByteDance | `Default` / `SVIP` | `SeeDance2` | — | Seedream images use `Default`; Seedance video uses `SeeDance2` |
| Alibaba | `Default` / `SVIP` | `Wan&HappyHorse` | — | Qwen text uses `Default`; Wan / HappyHorse video uses the dedicated group |
| DeepSeek / Moonshot / Zhipu / MiniMax | `Default` / `SVIP` | `ClaudeCode` (Anthropic-native compatible) | — | Official relay only; no separate reverse group |
| xAI | `Default` / `SVIP` | `CodexResponses` | — | Official relay only |
| Others (Meta / Mistral / Cohere / iFlytek / 01.AI / FLUX / MAI) | `Default` / `SVIP` | — | — | Official relay only |

## Choosing by key model

### Coding Agents

| Scenario | Recommended group | Notes |
| - | - | - |
| Claude Code + Claude Opus / Sonnet | `ClaudeCode` (daily) / `Claude_Enterprise` (production) | `ClaudeCode` is 95% and supports Anthropic native; production and enterprise workloads are more stable on `Claude_Enterprise` |
| Codex + GPT-5.6 / 6.x | `CodexResponses` / `Default` | Use `CodexResponses` when you need the Responses API; cost-sensitive users can test `Codex_Reverse` (non-production) |
| Gemini CLI + Gemini 3.x | `Default` | `Gemini_Reverse` covers up to 3.8 and is for learning / non-production |
| Domestic models + Claude Code | `ClaudeCode` | qwen / glm / kimi / deepseek / minimax and others compatible with `/v1/messages` |
| OpenClaw / Cherry and similar Agents | `Default`, or `CodexResponses` for Codex models | Prefer `Default` for mixed models; use `CodexResponses` for Codex-only workloads |

### Image models

For image generation, check three things first: do you need base64 or a URL; is this a peak/production workload that needs an enterprise fallback; and are you willing to use a cheaper reverse model. The default is `Default`; add `image2Enterprise` / `NB-Enterprise` for stability; use `image2_OSS` / `NB_OSS` when you need a URL; `-all` / `-vip` are lower-cost reverse models for cost-sensitive, non-hard-real-time workloads.

| Model | Default / production | Stable fallback | URL output | Notes |
| - | - | - | - | - |
| GPT Image 2.6 / 2.5 / 2 | `Default` | `image2Enterprise` (1.2×) | `image2_OSS` | `-all` is lower-cost; 2K/4K are upscaled; `-vip` is the Adobe reverse line |
| Nano Banana Pro / 2 / 2.1 | `Default` | `NB-Enterprise` (`Default` × 1.4) | `NB_OSS` | `NB_OSS` returns both a URL and Base64 |
| Seedream 5.5 / 5.0 / 4.5 | `Default` | — | — | Official relay |
| FLUX 2 / MAI Image 2.6 | `Default` | — | — | Official relay |

### Video models

| Model | Group | Notes |
| - | - | - |
| Seedance 2.0 / 2.5 | `SeeDance2` | 0.18 multiplier; 2.5, standard, fast, and mini all use this group |
| Wan 2.6 / 2.7 / 3.0, HappyHorse 1.0 / 1.1 | `Wan&HappyHorse` | 0.14 multiplier; billed per call or per second |

### Long context, multimodal, and retrieval

Long-context models, multimodal understanding, Embedding, Rerank, and TTS use `Default` by default; use `SVIP` when you need a scarce Pro model. These workloads do not have separate reverse groups.

## Group selection checklist

* **Select the group on the token first**: Console → Tokens → Group. Use a fallback group when the primary group fails.
* **A model outside the group returns an error**: The common message is "no available channel for this model". Check the model's supported groups before suspecting the key.
* **Reverse groups do not promise the same SLA as official relay**: Production, enterprise, and complex Agent workloads should prefer official relay. Reverse groups are for learning, cost-sensitive, and non-critical tasks.
* **The console is authoritative for prices**: Recharge bonuses follow the promotion rules. The "with 10% / 20% bonus" columns are effective prices, not an extra charge.
* **Start with `Default` if unsure**: Once the request works, switch to a dedicated group based on stability or cost.

> [!WARNING]
>
> `Claude_Reverse`, `Codex_Reverse`, and `Gemini_Reverse` are reverse-engineered resources. Availability fluctuates with the upstream resource pool, and they do not promise the same SLA as official relay. Production environments, enterprise customers, and complex Agent workflows should prefer official relay or enterprise groups.

## APIYI channel principles

| Principle | What we do |
| - | - |
| Official relay by default | APIYI uses vendor official APIs and official cloud marketplaces by default. Reverse is a cost supplement, not the default channel. |
| Label the source | `Claude_Reverse`, `Codex_Reverse`, and `Gemini_Reverse` are explicitly labeled as reverse; we never describe them as official direct connections. |
| Official relay for production | Production and enterprise workloads prefer `Default`, `Claude_Enterprise`, `CodexResponses`, `image2Enterprise`, and `NB-Enterprise`. |
| Official relay for complex Agents | Claude Code, Codex, Gemini CLI, and OpenClaw tool-calling and multi-turn tasks should default to official relay. |
| Where reverse fits | Personal learning, cost-sensitive, and non-critical tasks. Reverse does not promise the same SLA, concurrency, or feature completeness as official relay. |
| Enterprise guarantees | Use `Claude_Enterprise`, `image2Enterprise`, or `NB-Enterprise` when guarantees matter. Contact support for SLA and procurement options. |
| Console is authoritative | Upstream resources, model versions, and group multipliers can change. Trust the console and [Live Updates](https://docs.apiyi.com/en/live.md). |

## Related docs

> [card] [Official-Relay vs Reverse-Engineered APIs](https://docs.apiyi.com/en/faq/official-vs-reverse-api.md)
>
> Compare stability, protocol, caching, concurrency, and compliance

> [card] [Tokens & Groups](https://docs.apiyi.com/en/faq/token-and-groups.md)
>
> Create tokens and configure primary and fallback groups

> [card] [What Is a Group?](https://docs.apiyi.com/en/faq/groups-explained.md)
>
> User groups vs token groups

> [card] [Cache Billing](https://docs.apiyi.com/en/faq/cache-billing.md)
>
> Cache behavior and billing for Claude, OpenAI, and Gemini

> [card] [Model Pricing](https://docs.apiyi.com/en/models/index.md)
>
> Prices and specifications for every model

> [card] [Image and Video Models](https://docs.apiyi.com/en/api-capabilities/image-video-models.md)
>
> Generation guides and parameters by model


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.