@ai-sdk/xai 4.0.0-beta.7 → 4.0.0-beta.75

package/docs/01-xai.mdx

@@ -1,6 +1,6 @@
 ---
 title: xAI Grok
-description: Learn how to use xAI Grok.
+description: Learn how to use xAI Grok and Imagine.
 ---
 
 # xAI Grok Provider
@@ -73,10 +73,10 @@ You can use the following optional settings to customize the xAI provider instan
 ## Language Models
 
 You can create [xAI models](https://console.x.ai) using a provider instance. The
-first argument is the model id, e.g. `grok-3`.
+first argument is the model id, e.g. `grok-4.20-non-reasoning`.
 
 ```ts
-const model = xai('grok-3');
+const model = xai('grok-4.20-non-reasoning');
 ```
 
 By default, `xai(modelId)` uses the Responses API. To use the [Chat Completions API](https://docs.x.ai/docs/api-reference#chat-completions) (legacy), use `xai.chat(modelId)`.
@@ -90,7 +90,7 @@ import { xai } from '@ai-sdk/xai';
 import { generateText } from 'ai';
 
 const { text } = await generateText({
-  model: xai('grok-3'),
+  model: xai('grok-4.20-non-reasoning'),
   prompt: 'Write a vegetarian lasagna recipe for 4 people.',
 });
 ```
@@ -99,12 +99,76 @@ xAI language models can also be used in the `streamText` function
 and support structured data generation with [`Output`](/docs/reference/ai-sdk-core/output)
 (see [AI SDK Core](/docs/ai-sdk-core)).
 
+### Reasoning Effort
+
+For reasoning-capable models you can control how much effort the model spends
+thinking before responding via `providerOptions.xai.reasoningEffort`. This
+works for both the Responses API (default) and the Chat Completions API
+(`xai.chat()`).
+
+```ts
+import { xai } from '@ai-sdk/xai';
+import { generateText } from 'ai';
+
+const { text } = await generateText({
+  model: xai('grok-4.3'),
+  prompt: 'Explain quantum entanglement.',
+  providerOptions: {
+    xai: { reasoningEffort: 'medium' },
+  },
+});
+```
+
+Supported values:
+
+- `'none'` — Disables reasoning entirely; no thinking tokens are used. Best
+  for simple use cases that need a near-instant response. Supported by
+  `grok-4.3` and newer reasoning models.
+- `'low'` (default) — Uses some reasoning tokens, but still fast. Good for
+  general agentic use and tool calling.
+- `'medium'` — More thinking for less-latency-sensitive applications such as
+  complex data analysis and long-context reasoning.
+- `'high'` — More reasoning tokens for deeper thinking. Suited for very
+  challenging problems, complex math, multi-step logic, and competition-level
+  tasks.
+
+<Note>
+  Not every Grok model accepts every value. See xAI's [reasoning
+  docs](https://docs.x.ai/docs/guides/reasoning) for the values supported by
+  your selected model. `'none'` requires `grok-4.3` or newer.
+</Note>
+
+## Realtime Models
+
+<Note type="warning">Realtime is an experimental feature.</Note>
+
+You can create models that call the [xAI Realtime API](https://docs.x.ai/docs/guides/realtime)
+using the `.experimental_realtime()` factory method.
+
+```ts
+import { xai } from '@ai-sdk/xai';
+
+const model = xai.experimental_realtime('grok-voice-latest');
+```
+
+Realtime sessions run in the browser and require a short-lived token created on
+your server with `xai.experimental_realtime.getToken()`:
+
+```ts
+const token = await xai.experimental_realtime.getToken({
+  model: 'grok-voice-latest',
+});
+```
+
+See [Realtime](/docs/ai-sdk-core/realtime) for the complete setup and tool
+calling pattern.
+
 ## Responses API (Agentic Tools)
 
 The xAI Responses API is the default when using `xai(modelId)`. You can also use `xai.responses(modelId)` explicitly. This enables the model to autonomously orchestrate tool calls and research on xAI's servers.
 
 ```ts
-const model = xai.responses('grok-4-fast-non-reasoning');
+const model = xai.responses('grok-4.20-non-reasoning');
 ```
 
 The Responses API provides server-side tools that the model can autonomously execute during its reasoning process:
@@ -132,7 +196,11 @@ const { text } = await generateText({
       role: 'user',
       content: [
         { type: 'text', text: 'What do you see in this image?' },
-        { type: 'image', image: fs.readFileSync('./image.png') },
+        {
+          type: 'file',
+          mediaType: 'image',
+          data: fs.readFileSync('./image.png'),
+        },
       ],
     },
   ],
@@ -148,7 +216,7 @@ import { xai } from '@ai-sdk/xai';
 import { generateText } from 'ai';
 
 const { text, sources } = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt: 'What are the latest developments in AI?',
   tools: {
     web_search: xai.tools.webSearch({
@@ -172,9 +240,13 @@ console.log('Citations:', sources);
 
   Exclude specified domains from search (max 5). Cannot be used with `allowedDomains`.
 
+- **enableImageSearch** _boolean_
+
+  Allow the model to perform image search as a separate Web Search mode when images are useful for the answer. The model can choose between regular web search and image search; responses can include Markdown image embeds when relevant.
+
 - **enableImageUnderstanding** _boolean_
 
-  Enable the model to view and analyze images found during search. Increases token usage.
+  Enable the model to view and analyze images found during search. Increases token usage. To allow explicitly searching for images, use `enableImageSearch`.
 
 ### X Search Tool
 
@@ -182,7 +254,7 @@ The X search tool enables searching X (Twitter) for posts, with filtering by han
 
 ```ts
 const { text, sources } = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt: 'What are people saying about AI on X this week?',
   tools: {
     x_search: xai.tools.xSearch({
@@ -228,7 +300,7 @@ The code execution tool enables the model to write and execute Python code for c
 
 ```ts
 const { text } = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt:
     'Calculate the compound interest for $10,000 at 5% annually for 10 years',
   tools: {
@@ -243,7 +315,7 @@ The view image tool enables the model to view and analyze images:
 
 ```ts
 const { text } = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt: 'Describe what you see in the image',
   tools: {
     view_image: xai.tools.viewImage(),
@@ -257,7 +329,7 @@ The view X video tool enables the model to view and analyze videos from X (Twitt
 
 ```ts
 const { text } = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt: 'Summarize the content of this X video',
   tools: {
     view_x_video: xai.tools.viewXVideo(),
@@ -271,7 +343,7 @@ The MCP server tool enables the model to connect to remote [Model Context Protoc
 
 ```ts
 const { text } = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt: 'Use the weather tool to check conditions in San Francisco',
   tools: {
     weather_server: xai.tools.mcpServer({
@@ -319,7 +391,7 @@ import { xai, type XaiLanguageModelResponsesOptions } from '@ai-sdk/xai';
 import { streamText } from 'ai';
 
 const result = streamText({
-  model: xai.responses('grok-4-1-fast-reasoning'),
+  model: xai.responses('grok-4.20-reasoning'),
   prompt: 'What documents do you have access to?',
   tools: {
     file_search: xai.tools.fileSearch({
@@ -352,8 +424,8 @@ const result = streamText({
   Include file search results in the response. When set to `['file_search_call.results']`, the response will contain the actual search results with file content and scores.
 
 <Note>
-  File search requires grok-4 family models and the Responses API. Vector stores
-  can be created using the [xAI
+  File search requires grok-4 family models (including grok-4.20) and the
+  Responses API. Vector stores can be created using the [xAI
   API](https://docs.x.ai/docs/guides/using-collections/api).
 </Note>
 
@@ -365,8 +437,8 @@ You can combine multiple server-side tools for comprehensive research:
 import { xai } from '@ai-sdk/xai';
 import { streamText } from 'ai';
 
-const { fullStream } = streamText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+const { stream } = streamText({
+  model: xai.responses('grok-4.20-non-reasoning'),
   prompt: 'Research AI safety developments and calculate risk metrics',
   tools: {
     web_search: xai.tools.webSearch(),
@@ -382,7 +454,7 @@ const { fullStream } = streamText({
   },
 });
 
-for await (const part of fullStream) {
+for await (const part of stream) {
   if (part.type === 'text-delta') {
     process.stdout.write(part.text);
   } else if (part.type === 'source' && part.sourceType === 'url') {
@@ -400,7 +472,7 @@ import { xai, type XaiLanguageModelResponsesOptions } from '@ai-sdk/xai';
 import { generateText } from 'ai';
 
 const result = await generateText({
-  model: xai.responses('grok-4-fast-non-reasoning'),
+  model: xai.responses('grok-4.20-non-reasoning'),
   providerOptions: {
     xai: {
       reasoningEffort: 'high',
@@ -412,9 +484,9 @@ const result = await generateText({
 
 The following provider options are available:
 
-- **reasoningEffort** _'low' | 'medium' | 'high'_
+- **reasoningEffort** _'none' | 'low' | 'medium' | 'high'_
 
-  Control the reasoning effort for the model. Higher effort may produce more thorough results at the cost of increased latency and token usage.
+  Control the reasoning effort for the model. See [Reasoning Effort](#reasoning-effort) for details on each value. `'none'` disables reasoning entirely (requires `grok-4.3` or newer).
 
 - **logprobs** _boolean_
 
@@ -445,19 +517,16 @@ The following provider options are available:
 
 | Model                         | Image Input         | Object Generation   | Tool Usage          | Tool Streaming      | Reasoning           |
 | ----------------------------- | ------------------- | ------------------- | ------------------- | ------------------- | ------------------- |
-| `grok-4-1`                    | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
+| `grok-4.20-reasoning`         | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+| `grok-4.20-non-reasoning`     | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
 | `grok-4-1-fast-reasoning`     | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | `grok-4-1-fast-non-reasoning` | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
-| `grok-4-fast-non-reasoning`   | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
+| `grok-4-1`                    | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
 | `grok-4-fast-reasoning`       | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+| `grok-4-fast-non-reasoning`   | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
 | `grok-code-fast-1`            | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
-| `grok-4`                      | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
-| `grok-4-0709`                 | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
-| `grok-4-latest`               | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
 | `grok-3`                      | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
-| `grok-3-latest`               | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Cross size={18} /> |
 | `grok-3-mini`                 | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
-| `grok-3-mini-latest`          | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 
 <Note>
   The table above lists popular models. Please see the [xAI
@@ -465,6 +534,198 @@ The following provider options are available:
   can also pass any available provider model ID as a string if needed.
 </Note>
 
+## Speech Models
+
+You can create models that call the [xAI Text to Speech API](https://docs.x.ai/developers/model-capabilities/audio/text-to-speech)
+using the `.speech()` factory method. xAI's text-to-speech endpoint does not
+require a model identifier.
+
+```ts
+const model = xai.speech();
+```
+
+Use xAI speech models with the `generateSpeech` function:
+
+```ts
+import { xai } from '@ai-sdk/xai';
+import { generateSpeech } from 'ai';
+
+const result = await generateSpeech({
+  model: xai.speech(),
+  text: 'Hello from the AI SDK!',
+  voice: 'ara',
+  language: 'en',
+  outputFormat: 'mp3',
+  speed: 1.1,
+});
+```
+
+### Supported Parameters
+
+- **text** _string_ (required)
+
+  Text to synthesize. You can include xAI speech tags such as `[pause]`,
+  `[laugh]`, or `<whisper>...</whisper>` directly in the text.
+
+- **voice** _string_
+
+  The xAI voice ID. Defaults to `'eve'`. Built-in voice IDs are `'eve'`,
+  `'ara'`, `'rex'`, `'sal'`, and `'leo'`; custom voice IDs are also accepted.
+
+- **language** _string_
+
+  A BCP-47 language code or `'auto'` for automatic detection. Defaults to
+  `'auto'` when omitted.
+
+- **speed** _number_
+
+  Speech rate multiplier from `0.7` to `1.5`.
+
+- **outputFormat** _string_
+
+  The audio codec to generate. Supported values are `'mp3'`, `'wav'`, `'pcm'`,
+  `'mulaw'`, and `'alaw'`. Defaults to `'mp3'`.
+
+<Note>
+  xAI does not expose a separate `instructions` field for text-to-speech. Use
+  speech tags in `text` to control expressive delivery.
+</Note>
+
+### Provider Options
+
+You can pass xAI-specific controls using `providerOptions.xai`:
+
+```ts
+import { xai, type XaiSpeechModelOptions } from '@ai-sdk/xai';
+import { generateSpeech } from 'ai';
+
+const result = await generateSpeech({
+  model: xai.speech(),
+  text: 'A high fidelity narration sample.',
+  outputFormat: 'mp3',
+  providerOptions: {
+    xai: {
+      sampleRate: 44100,
+      bitRate: 192000,
+      optimizeStreamingLatency: 1,
+      textNormalization: true,
+    } satisfies XaiSpeechModelOptions,
+  },
+});
+```
+
+- **sampleRate** _8000 | 16000 | 22050 | 24000 | 44100 | 48000_
+
+  Sample rate of the output audio in Hz.
+
+- **bitRate** _32000 | 64000 | 96000 | 128000 | 192000_
+
+  MP3 bit rate in bits per second. Only applies when `outputFormat` is
+  `'mp3'`.
+
+- **optimizeStreamingLatency** _0 | 1 | 2_
+
+  Latency optimization level. Higher values reduce time to first audio with a
+  quality tradeoff at chunk boundaries.
+
+- **textNormalization** _boolean_
+
+  Whether to normalize written-form input text before synthesizing speech.
+
+### Model Capabilities
+
+| Model     | Language            | Speed               | Output Formats             |
+| --------- | ------------------- | ------------------- | -------------------------- |
+| `default` | <Check size={18} /> | <Check size={18} /> | mp3, wav, pcm, mulaw, alaw |
+
+## Transcription Models
+
+You can create models that call the [xAI Speech to Text API](https://docs.x.ai/developers/model-capabilities/audio/speech-to-text)
+using the `.transcription()` factory method. xAI's batch transcription endpoint
+does not require a model identifier.
+
+```ts
+const model = xai.transcription();
+```
+
+Use xAI transcription models with the `transcribe` function:
+
+```ts
+import { xai } from '@ai-sdk/xai';
+import { transcribe } from 'ai';
+import { readFile } from 'fs/promises';
+
+const result = await transcribe({
+  model: xai.transcription(),
+  audio: await readFile('meeting.mp3'),
+});
+```
+
+### Provider Options
+
+You can pass xAI-specific controls using `providerOptions.xai`:
+
+```ts
+import { xai, type XaiTranscriptionModelOptions } from '@ai-sdk/xai';
+import { transcribe } from 'ai';
+import { readFile } from 'fs/promises';
+
+const result = await transcribe({
+  model: xai.transcription(),
+  audio: await readFile('meeting.mp3'),
+  providerOptions: {
+    xai: {
+      language: 'en',
+      format: true,
+      keyterm: ['AI SDK', 'Grok'],
+      diarize: true,
+    } satisfies XaiTranscriptionModelOptions,
+  },
+});
+```
+
+- **audioFormat** _`pcm` | `mulaw` | `alaw`_
+
+  Audio encoding for raw, headerless input audio.
+
+- **sampleRate** _8000 | 16000 | 22050 | 24000 | 44100 | 48000_
+
+  Sample rate of the input audio in Hz.
+
+- **language** _string_
+
+  Language code used for inverse text normalization.
+
+- **format** _boolean_
+
+  Enables inverse text normalization. Requires `language`.
+
+- **multichannel** _boolean_
+
+  Enables per-channel transcription for interleaved multichannel audio.
+
+- **channels** _2 | 3 | 4 | 5 | 6 | 7 | 8_
+
+  Number of interleaved audio channels.
+
+- **diarize** _boolean_
+
+  Enables speaker diarization.
+
+- **keyterm** _string | string[]_
+
+  One or more terms to bias transcription toward.
+
+- **fillerWords** _boolean_
+
+  Includes filler words such as `uh` and `um` in the transcript.
+
+### Model Capabilities
+
+| Model     | Word Timestamps     | Diarization         | Multichannel        |
+| --------- | ------------------- | ------------------- | ------------------- |
+| `default` | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+
 ## Image Models
 
 You can create xAI image models using the `.image()` factory method. For more on image generation with the AI SDK see [generateImage()](/docs/reference/ai-sdk-core/generate-image).
@@ -516,7 +777,7 @@ const { images } = await generateImage({
 
 #### Multi-Image Editing
 
-Combine or reference multiple input images (up to 3) in the prompt:
+Combine or reference multiple input images in the prompt:
 
 ```ts
 import { xai } from '@ai-sdk/xai';
@@ -554,37 +815,53 @@ const { images } = await generateImage({
 
 <Note>
   Input images can be provided as `Buffer`, `ArrayBuffer`, `Uint8Array`, or
-  base64-encoded strings. Up to 3 input images are supported per request.
+  base64-encoded strings.
 </Note>
 
-### Model-specific options
+### Image Provider Options
 
-You can customize the image generation behavior with model-specific settings:
+You can customize the image generation behavior with provider-specific settings via `providerOptions.xai`:
 
 ```ts
-import { xai } from '@ai-sdk/xai';
+import { xai, type XaiImageModelOptions } from '@ai-sdk/xai';
 import { generateImage } from 'ai';
 
 const { images } = await generateImage({
-  model: xai.image('grok-imagine-image'),
+  model: xai.image('grok-imagine-image-pro'),
   prompt: 'A futuristic cityscape at sunset',
   aspectRatio: '16:9',
-  n: 2,
+  providerOptions: {
+    xai: {
+      resolution: '2k',
+      quality: 'high',
+    } satisfies XaiImageModelOptions,
+  },
 });
 ```
 
-### Model Capabilities
+- **resolution** _'1k' | '2k'_
+
+  Output resolution. `1k` produces ~1024×1024 images, `2k` produces ~2048×2048
+  images (actual dimensions vary based on aspect ratio). Available for
+  `grok-imagine-image-pro`.
+
+- **quality** _'low' | 'medium' | 'high'_
 
-| Model                | Aspect Ratios                                                                                               | Image Editing       |
-| -------------------- | ----------------------------------------------------------------------------------------------------------- | ------------------- |
-| `grok-imagine-image` | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3`, `2:1`, `1:2`, `19.5:9`, `9:19.5`, `20:9`, `9:20`, `auto` | <Check size={18} /> |
+  Image quality level. Higher quality may increase generation time.
+
+### Image Model Capabilities
+
+| Model                    | Resolution | Aspect Ratios                                                                                               | Image Editing       |
+| ------------------------ | ---------- | ----------------------------------------------------------------------------------------------------------- | ------------------- |
+| `grok-imagine-image-pro` | `1k`, `2k` | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3`, `2:1`, `1:2`, `19.5:9`, `9:19.5`, `20:9`, `9:20`, `auto` | <Check size={18} /> |
+| `grok-imagine-image`     | `1k`       | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3`, `2:1`, `1:2`, `19.5:9`, `9:19.5`, `20:9`, `9:20`, `auto` | <Check size={18} /> |
 
 ## Video Models
 
 You can create xAI video models using the `.video()` factory method.
 For more on video generation with the AI SDK see [generateVideo()](/docs/reference/ai-sdk-core/generate-video).
 
-This provider supports three video generation modes: text-to-video, image-to-video, and video editing.
+This provider supports standard video generation from text prompts or image input, plus explicit video editing, video extension, and reference-to-video (R2V) operations.
 
 ### Text-to-Video
 
@@ -594,7 +871,7 @@ Generate videos from text prompts:
 import { xai, type XaiVideoModelOptions } from '@ai-sdk/xai';
 import { experimental_generateVideo as generateVideo } from 'ai';
 
-const { videos } = await generateVideo({
+const { video } = await generateVideo({
   model: xai.video('grok-imagine-video'),
   prompt: 'A chicken flying into the sunset in the style of 90s anime.',
   aspectRatio: '16:9',
@@ -607,15 +884,15 @@ const { videos } = await generateVideo({
 });
 ```
 
-### Image-to-Video
+### Generation with Image Input
 
-Generate videos using an image as the starting frame with an optional text prompt:
+Generate videos using an image as the starting frame with an optional text prompt. This uses the standard generation path rather than a separate provider mode:
 
 ```ts
 import { xai, type XaiVideoModelOptions } from '@ai-sdk/xai';
 import { experimental_generateVideo as generateVideo } from 'ai';
 
-const { videos } = await generateVideo({
+const { video } = await generateVideo({
   model: xai.video('grok-imagine-video'),
   prompt: {
     image: 'https://example.com/start-frame.png',
@@ -638,11 +915,12 @@ Edit an existing video using a text prompt by providing a source video URL via p
 import { xai, type XaiVideoModelOptions } from '@ai-sdk/xai';
 import { experimental_generateVideo as generateVideo } from 'ai';
 
-const { videos } = await generateVideo({
+const { video } = await generateVideo({
   model: xai.video('grok-imagine-video'),
   prompt: 'Give the person sunglasses and a hat',
   providerOptions: {
     xai: {
+      mode: 'edit-video',
       videoUrl: 'https://example.com/source-video.mp4',
       pollTimeoutMs: 600000, // 10 minutes
     } satisfies XaiVideoModelOptions,
@@ -668,6 +946,7 @@ import { experimental_generateVideo as generateVideo } from 'ai';
 
 const providerOptions = {
   xai: {
+    mode: 'edit-video',
     videoUrl: 'https://example.com/source-video.mp4',
     pollTimeoutMs: 600000,
   } satisfies XaiVideoModelOptions,
@@ -689,19 +968,107 @@ const [withSunglasses, withScarf] = await Promise.all([
     model: xai.video('grok-imagine-video'),
     prompt: 'Add sunglasses',
     providerOptions: {
-      xai: { videoUrl: step1VideoUrl, pollTimeoutMs: 600000 },
+      xai: {
+        mode: 'edit-video',
+        videoUrl: step1VideoUrl,
+        pollTimeoutMs: 600000,
+      },
     },
   }),
   generateVideo({
     model: xai.video('grok-imagine-video'),
     prompt: 'Add a scarf',
     providerOptions: {
-      xai: { videoUrl: step1VideoUrl, pollTimeoutMs: 600000 },
+      xai: {
+        mode: 'edit-video',
+        videoUrl: step1VideoUrl,
+        pollTimeoutMs: 600000,
+      },
     },
   }),
 ]);
 ```
 
+### Video Extension
+
+Extend an existing video from its last frame. The `duration` controls the length of the extension only, not the total output. The output inherits `aspectRatio` and `resolution` from the source video.
+
+```ts
+import { xai, type XaiVideoModelOptions } from '@ai-sdk/xai';
+import { experimental_generateVideo as generateVideo } from 'ai';
+
+// Step 1: Generate a source video
+const source = await generateVideo({
+  model: xai.video('grok-imagine-video'),
+  prompt: 'A cat sitting on a sunlit windowsill, tail gently swishing.',
+  duration: 5,
+  aspectRatio: '16:9',
+  providerOptions: {
+    xai: {
+      pollTimeoutMs: 600000,
+    } satisfies XaiVideoModelOptions,
+  },
+});
+
+const sourceUrl = source.providerMetadata?.xai?.videoUrl as string;
+
+// Step 2: Extend the video with a new scene
+const extended = await generateVideo({
+  model: xai.video('grok-imagine-video'),
+  prompt: 'The cat turns its head, notices a butterfly, and leaps off.',
+  duration: 6,
+  providerOptions: {
+    xai: {
+      mode: 'extend-video',
+      videoUrl: sourceUrl,
+      pollTimeoutMs: 600000,
+    } satisfies XaiVideoModelOptions,
+  },
+});
+```
+
+<Note>
+  Video extension does not support custom `aspectRatio` or `resolution` — the
+  output inherits those from the source video. `duration` is supported and
+  controls how long the extension is (not the total video length).
+</Note>
+
+### Reference-to-Video (R2V)
+
+Provide reference images to guide the video's style and content. Unlike image-to-video, reference images are not used as the first frame — the model incorporates their visual elements into the generated video. Each reference image can be a public HTTPS URL or a base64 data URI.
+
+```ts
+import { xai, type XaiVideoModelOptions } from '@ai-sdk/xai';
+import { experimental_generateVideo as generateVideo } from 'ai';
+
+const { video } = await generateVideo({
+  model: xai.video('grok-imagine-video'),
+  prompt:
+    'The comic cat from <IMAGE_1> and the comic dog from <IMAGE_2> ' +
+    'are having a playful chase through a sunlit park. ' +
+    'Cinematic slow-motion, warm afternoon light.',
+  duration: 8,
+  aspectRatio: '16:9',
+  providerOptions: {
+    xai: {
+      mode: 'reference-to-video',
+      referenceImageUrls: [
+        'https://example.com/comic-cat.png',
+        'https://example.com/comic-dog.png',
+      ],
+      pollTimeoutMs: 600000,
+    } satisfies XaiVideoModelOptions,
+  },
+});
+```
+
+Use `<IMAGE_1>`, `<IMAGE_2>`, etc. in your prompt to reference specific images. Up to 7 reference images are supported per request.
+
+<Note>
+  Reference-to-video supports `duration`, `aspectRatio`, and `resolution`. Use
+  `mode` to select the operation — each mode is mutually exclusive.
+</Note>
+
 ### Video Provider Options
 
 The following provider options are available via `providerOptions.xai`.
@@ -721,10 +1088,27 @@ You can validate the provider options using the `XaiVideoModelOptions` type.
   `1280x720` maps to `720p` and `854x480` maps to `480p`.
   Use this provider option to pass the native format directly.
 
+- **mode** _'edit-video' | 'extend-video' | 'reference-to-video'_
+
+  Selects the explicit video operation. Each mode is mutually exclusive:
+  - `'edit-video'` — edit an existing video (requires `videoUrl`)
+  - `'extend-video'` — extend a video from its last frame (requires `videoUrl`)
+  - `'reference-to-video'` — generate from reference images (requires `referenceImageUrls`)
+
+  When omitted, standard generation is used. Legacy inputs are still auto-detected from fields for backward compatibility.
+
 - **videoUrl** _string_
 
-  URL of a source video for video editing. When provided, the prompt is used
-  to describe the desired edits to the video.
+  URL of a source video. Used with `mode: 'edit-video'` for video editing
+  and `mode: 'extend-video'` for video extension.
+
+- **referenceImageUrls** _string[]_
+
+  Array of reference image URLs (1–7 images) or base64 data URIs for
+  reference-to-video (R2V) generation. The model incorporates visual
+  elements from these images without using them as the first frame. Use
+  `<IMAGE_1>`, `<IMAGE_2>`, etc. in the prompt to reference specific
+  images. Used with `mode: 'reference-to-video'`.
 
 <Note>
   Video generation is an asynchronous process that can take several minutes.
@@ -744,14 +1128,21 @@ desired ratio.
 
 For **video editing**, the output matches the input video's aspect ratio and
 resolution. Custom `duration`, `aspectRatio`, and `resolution` are not
-supported - the output resolution is capped at 720p (e.g., a 1080p input
+supported — the output resolution is capped at 720p (e.g., a 1080p input
 will be downsized to 720p).
 
+For **video extension**, the output inherits `aspectRatio` and `resolution`
+from the source video. `duration` is supported and controls only the
+extension length.
+
+For **reference-to-video (R2V)**, you can specify `duration`, `aspectRatio`,
+and `resolution` just like text-to-video.
+
 ### Video Model Capabilities
 
-| Model                | Duration | Aspect Ratios                                     | Resolution     | Image-to-Video      | Video Editing       |
-| -------------------- | -------- | ------------------------------------------------- | -------------- | ------------------- | ------------------- |
-| `grok-imagine-video` | 1–15s    | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3` | `480p`, `720p` | <Check size={18} /> | <Check size={18} /> |
+| Model                | Duration | Aspect Ratios                                     | Resolution     | Image-to-Video      | Editing             | Extension           | R2V                 |
+| -------------------- | -------- | ------------------------------------------------- | -------------- | ------------------- | ------------------- | ------------------- | ------------------- |
+| `grok-imagine-video` | 1–15s    | `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3` | `480p`, `720p` | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 
 <Note>
   You can also pass any available provider model ID as a string if needed.