Proposal: Add capabilities field to /v1/models response

Problem

Every LLM framework (LangChain, LiteLLM, Vercel AI SDK, etc.) maintains its own hardcoded model registry to answer basic questions like:

• Does this model support tool/function calling?
• Does it support vision (image inputs)?
• Does it support structured output / JSON mode?
• What is the context window size?

These registries are always stale, always incomplete, and always divergent. Every new model release triggers a wave of PRs across dozens of repos just to update capability flags.

Root Cause

The /v1/models endpoint returns almost no useful metadata:

{
  "id": "gpt-4o",
  "object": "model",
  "created": 1715367049,
  "owned_by": "system"
}

There is no machine-readable way to discover what a model can actually do.

Proposal

Add a capabilities object to the /v1/models/{model} response:

{
  "id": "gpt-4o",
  "object": "model",
  "created": 1715367049,
  "owned_by": "system",
  "capabilities": {
    "tool_calling": true,
    "vision": true,
    "structured_output": true,
    "json_mode": true,
    "streaming": true,
    "context_window": 128000,
    "max_output_tokens": 16384
  }
}

This is a backwards-compatible, additive change. Existing clients would simply ignore the new field.

Impact

• Eliminates the need for hardcoded model registries across the ecosystem
• Enables dynamic capability negotiation (agents can adapt to whatever model they are given)
• Reduces maintenance burden for every framework that builds on the OpenAI-compatible API
• Sets a standard that other providers (Anthropic, Google, Mistral, etc.) can adopt

Prior Art

• Blog post with detailed analysis: https://h3manth.com/scribe/model-capability-discovery/
• Hardware and browser APIs have long exposed capability discovery (e.g., navigator.mediaDevices, OpenGL extensions)

This would be a small schema change with outsized ecosystem impact.

Start by locating the OpenAPI schema for the /v1/models endpoint and compare the collection and per-model response shapes. Done means the schema defines the proposed capabilities object and its listed fields while preserving the existing response fields and documenting the additive behavior.