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

# Capabilities

> What each chat model can do — chat, streaming, vision, PDF, tools, JSON mode.

Each model declares capabilities as boolean flags. You can read them via `GET /v1/models/{slug}` to branch feature-by-feature in your app.

| Capability          | Meaning                                                                      |
| ------------------- | ---------------------------------------------------------------------------- |
| `supportsChat`      | Accepts chat completions format                                              |
| `supportsStreaming` | Can stream tokens via SSE                                                    |
| `supportsVision`    | Accepts image inputs                                                         |
| `supportsPdf`       | Accepts PDF natively (otherwise gateway auto-converts)                       |
| `supportsTools`     | Honours `tools` array + `tool_choice`                                        |
| `supportsJsonMode`  | `response_format: {type: "json_object"}` or JSON schema                      |
| `supportsImages`    | Generates images as output (different from Vision)                           |
| `isReasoning`       | Chain-of-thought reasoning model (slower but more accurate on complex tasks) |
| `isFlagship`        | Provider's top-of-line model (highest quality + cost)                        |

## Provider-by-provider highlights

### OpenAI

* **gpt-5-4** — flagship, all capabilities
* **gpt-4o** — vision ✓, tools ✓, JSON ✓, streaming ✓
* **o3 / o1** — reasoning ✓ (no streaming tokens — final result only)
* **gpt-4o-audio** — audio input natively in chat

### Anthropic

* **claude-opus-4-7** — flagship, vision ✓, PDF ✓, tools ✓
* **claude-opus-4-6** — vision ✓, PDF ✓, tools ✓
* **claude-sonnet-4-6** — same with better cost/perf
* **Audio input** — **not supported**; use `/v1/audio/transcriptions` first

### Google

* **gemini-2-5-pro** — vision ✓, PDF ✓, 2M context window
* **gemini-2-5-flash** — fast, huge context, multimodal
* **gemini-2-0-flash** — budget option

### xAI

* **grok-3** — vision ✓, JSON ✓

### Alibaba

* **qwen3-max** — vision + PDF + long context (262k)
* **qwen3-5-plus** — 1M context
* **qwen-vl-max** — vision-specialised

## PDF fallback

If a model without `supportsPdf` receives a PDF, our gateway:

1. Extracts text via pdftotext
2. Renders pages to PNG (for vision models)
3. Injects text + images into the message
4. Bills a small `pdf_extraction_per_page` fee

You don't lift a finger — works on Grok, DeepSeek, older models. The response behaves as if the model read the PDF natively.