Anvia

Models and media ​

Completion models ​

ts
const responses = new GrokClient({ apiKey })
    .completionModel({
    modelId: 'grok-4.5',
    api: 'responses',
})

const chat = new GrokClient({ apiKey }).completionModel({
    modelId: 'grok-4.5',
    api: 'chat',
})

Choose Responses for provider tools and native source normalization. Choose Chat only for workflows that require that endpoint and do not use provider-executed tools.

grok-4.5 remains a supported ID for coding, agentic tasks, and knowledge work, and newer IDs are available: grok-4.6, grok-4.6-latest, the grok-4.20 variants (grok-4.20, grok-4.20-non-reasoning, grok-4.20-0309-non-reasoning, grok-4.20-0309-reasoning, grok-4.20-multi-agent-0309), grok-4.3 and grok-4.3-latest, and grok-build-0.1. All support both Responses and Chat Completions; the v1 examples prefer Responses for web search, X search, code execution, and other provider tools.

Image generation ​

ts
const result = await grok.imageGenerationModel({ modelId: 'grok-imagine-image-quality' }).imageGeneration({
  prompt: 'A robot reviewing a pull request in a glass office',
  width: 1024,
  height: 1024,
})

The adapter supports the documented ratio set, maps 13:6 to 19.5:9, maps 6:13 to 9:19.5, and sends auto for other ratios. Invalid, non-finite, or non-positive dimensions throw.

Provider base64 images are decoded directly. URL images are fetched and returned as bytes; a URL response without fetch is an error.

Text-to-speech ​

ts
const speech = await grok.speechGenerationModel().speechGeneration({
  text: 'The run is complete.',
  voice: 'eve',
  speed: 1,
  providerOptions: {
    language: 'en',
    output_format: { codec: 'mp3', sample_rate: 24_000 },
  },
})

Speed must be exactly 1. Additional parameters must be an object.

Transcription ​

ts
const result = await grok.transcriptionModel().transcription({
  data: speech.audio.data,
  filename: 'speech.mp3',
  language: 'en',
})

The adapter uploads multipart bytes to the STT endpoint and retains provider JSON as rawResponse. A response without recognized text is rejected. Anvia prompt and temperature transcription options are rejected because this adapter does not map them to xAI.

Model listing ​

listModels() normalizes inventory but does not classify each model into completion or media types. Keep capability-specific allowlists in application configuration.

Built for Anvia.