Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/phpaisdk/openai/llms.txt

Use this file to discover all available pages before exploring further.

The aisdk/openai package ships a resources/models.json catalog that maps model IDs and ID patterns to their capabilities and limits. This catalog is queried at runtime to validate capability requests before making API calls — if a model handle is asked to perform an operation it does not support, the SDK throws a CapabilityNotSupportedException before a single HTTP byte is sent. The catalog is structured around two modalities — text for chat-completion models and image for generation models — with both exact IDs and wildcard patterns so that newly released variants of a model family are automatically covered.

​
Text Models

Text models are accessed via OpenAI::model($modelId) and drive the ->generate() and ->stream() paths through OpenAI’s /chat/completions endpoint. The table below lists every entry with "modalities": ["text"] from models.json, including the capabilities declared for each model and its token limits where specified.
Model / PatternStatusCapabilitiesContext TokensOutput Tokens
openai/gpt-oss-20bstabletext_generation, streaming, tool_calling, structured_output, text_input131 072—
gpt-4ostabletext_generation, streaming, tool_calling, structured_output, text_input, image_input128 00016 384
gpt-4o*stabletext_generation, streaming, tool_calling, structured_output, text_input, image_input——
gpt-4.1*stabletext_generation, streaming, tool_calling, structured_output, text_input, image_input, file_input——
gpt-*-audio*stabletext_generation, streaming, tool_calling, structured_output, text_input, audio_input——
o*stabletext_generation, streaming, tool_calling, structured_output, reasoning, text_input, image_input——
gpt-5*stabletext_generation, streaming, tool_calling, structured_output, reasoning, text_input, image_input, file_input——
To check whether a specific model ID supports a capability without making an API call, use OpenAI::model('gpt-4o')->supports(Capability::ImageInput).

​
Image Models

Image models are accessed via OpenAI::image($modelId) and use OpenAI’s /images/generations endpoint. All three image models in the catalog carry a single capability — image_generation — and are stable.
ModelStatusCapabilities
gpt-image-1stableimage_generation
dall-e-3stableimage_generation
dall-e-2stableimage_generation
use AiSdk\Generate;
use AiSdk\OpenAI;

$result = Generate::image()
    ->model(OpenAI::image('dall-e-3'))
    ->prompt('A photorealistic mountain landscape at sunrise')
    ->size('1024x1024')
    ->run();
The portable seed() option is not supported by OpenAI’s image API. Passing seed() on an image request will throw an InvalidArgumentException. If the specific model or endpoint does support seeding, use providerOptions('openai', ['raw' => ['seed' => 42]]) as an escape hatch.

​
Wildcard Matching

Model IDs ending in * are glob-style patterns. A pattern like gpt-4o* matches any model ID that starts with gpt-4o — for example, gpt-4o-mini, gpt-4o-2024-11-20, and gpt-4o-search-preview. Similarly, o* matches o1, o3, o3-mini, and future o-series releases. When the SDK resolves a model ID against the catalog it looks for an exact match first, then falls through to the most specific matching wildcard pattern. This means the explicit gpt-4o entry (with its precise context and output token limits) takes precedence over the gpt-4o* wildcard entry for that exact model ID, while gpt-4o-mini is resolved by the wildcard.
Pattern specificity is determined by the length and position of the literal characters before the *. More specific (longer literal prefix) patterns win over shorter ones.

​
Unknown Models

If a model ID does not match any entry — exact or wildcard — in the catalog, the SDK applies a minimal fallback policy rather than blocking the request outright:
  • TextGeneration is always allowed for unknown text model handles. The capability() method returns a CapabilitySupport with source = 'unknown-model-fallback', so you can detect this programmatically.
  • All other capabilities (ToolCalling, StructuredOutput, ImageInput, etc.) are reported as NotSupported and will trigger CapabilityNotSupportedException if the SDK tries to use them.
To opt in to additional capabilities on an unknown or preview model, call ->assume() or ->allowUnknownCapabilities() on the model handle: 
use AiSdk\Capability;
use AiSdk\OpenAI;

// Opt in to specific capabilities for a model not yet in the catalog
$model = OpenAI::model('gpt-4.2-preview')->assume([
    Capability::ToolCalling,
    Capability::StructuredOutput,
]);

// Opt in to all capabilities — useful during early-access / beta
$model = OpenAI::model('gpt-4.2-preview')->allowUnknownCapabilities();
->allowUnknownCapabilities() bypasses all catalog checks. The OpenAI API will still return a provider-level error if the model does not actually support the requested feature — but the SDK will not prevent the HTTP call from being made.

Build docs developers (and LLMs) love

Get started for freeTalk to us