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

# Features

> An overview of OpenPAI's core capabilities and differentiators

While aggregating all the major LLMs behind one interface, OpenPAI also provides the operations, billing and observability capabilities an enterprise-grade gateway needs.

## Unified access

<CardGroup cols={2}>
  <Card title="OpenAI-compatible protocol" icon="o">
    Fully compatible with `/v1/chat/completions`, `/v1/embeddings`, `/v1/images/generations`, `/v1/audio/*`, `/v1/responses`, `/v1/realtime` and more.
  </Card>

  <Card title="Claude native protocol" icon="a">
    Provides the `/v1/messages` endpoint with support for system prompts, tool use, vision input and thinking mode.
  </Card>

  <Card title="Gemini native protocol" icon="g">
    Provides the `/v1beta/models/{model}:generateContent` endpoint with full support for multimodal input and function calling.
  </Card>

  <Card title="Format conversion" icon="arrows-rotate">
    Automatic conversion between OpenAI ⇄ Claude / OpenAI → Gemini / Gemini → OpenAI formats.
  </Card>
</CardGroup>

## Model coverage

| Category            | Representative models                                                                                                     |
| ------------------- | ------------------------------------------------------------------------------------------------------------------------- |
| Chat                | GPT-5.5 / 5.4, o4-mini, Claude Opus 4.8, Claude Sonnet 4.6, Gemini 3.1 Pro / 3.5 Flash, DeepSeek-V3, Qwen3, GLM-4.6, etc. |
| Reasoning           | `gpt-5.x` (`reasoning_effort`), `o3` / `o4-mini`, Claude `*-thinking` / Opus `effort`, `gemini-3.1-pro-preview`           |
| Vision              | GPT-5.x, Claude 4.x, Gemini 3.x, Qwen-VL                                                                                  |
| Image generation    | `gpt-image-2`, DALL·E 3, Midjourney (Proxy / Plus), Nano Banana                                                           |
| Audio               | Whisper, TTS, gpt-realtime multimodal                                                                                     |
| Video               | Sora, Veo, Kling, etc.                                                                                                    |
| Embeddings / rerank | `text-embedding-3-*`, `bge-*`, Cohere, Jina Rerank                                                                        |
| Music               | Suno, Udio                                                                                                                |

See the full model list in the [Models overview](/en/models/overview).

## Billing & quota

* **Pay-as-you-go**: tokens are converted at the upstream's official multiplier, with lower rates on cache hits.
* **Per-token quota**: set a spending limit, expiry and model scope individually for each API key.
* **Top-ups & invoices**: manual top-ups assisted by a dedicated agent, with invoices on request.

See [Billing](/en/pricing/billing) for details.

## High-availability routing

<CardGroup cols={2}>
  <Card title="Multi-channel load" icon="scale-balanced">
    The same model can be bound to multiple upstream channels and split by weight.
  </Card>

  <Card title="Automatic retries" icon="rotate-right">
    Upstream errors automatically fail over to a healthy channel, transparently to the client.
  </Card>

  <Card title="Rate limiting" icon="gauge-high">
    Per-user / per-model RPM and TPM limits protect backend stability.
  </Card>

  <Card title="Real-time monitoring" icon="chart-line">
    The console offers dashboards for request volume, spend and error rate.
  </Card>
</CardGroup>

## Advanced capabilities

* **Chain of thought**: supports OpenAI `reasoning_effort`, Claude `thinking` and Gemini `thinking_budget`.
* **Function calling**: OpenAI Tools / Claude Tool Use / Gemini Function Calling all interoperate across protocols.
* **Structured output**: strictly constrain output with a `response_format` JSON Schema.
* **Multimodal input**: images (URL / base64), PDF, audio chunks and more.
* **Streaming responses**: SSE streams (Server-Sent Events), fully compatible with the OpenAI streaming protocol.

## Accounts & logs

* Sign in with email / GitHub / Discord / Telegram / LinuxDO / OIDC.
* 2FA / WebAuthn / Passkey two-factor authentication.
* Prompt and response content is not recorded — the console only lets you review each call's model, token usage and billing details.
* Self-hosted deployment is supported.