Capabilities

Every model is reached through the same three endpoints with the same key. The two tables show where the models differ. A dash means the model answers without that capability.

Shannon models

Hosted open-weight models

The id of a hosted model is its name and its variant joined by a hyphen, for example DeepSeek-V4-Pro-0813-3BIT-REAP.

What each column means

• Tools. You describe functions in tools, the model replies with the calls it wants made, and your code runs them. Every model takes tools. A forced tool_choice is applied by the hosted open-weight models. Function Calling
• Image input. The model reads images sent inside a user message, as a data URL or as an address. shannon-2-lite and shannon-2-pro read images in requests that also send tools or response_format. The Shannon 3 family reads documents as well. Images and files
• Reasoning. The reply can carry the reasoning text of the model next to the answer, in a field of its own. Reasoning effort
• Reasoning effort. A request field sets how deeply the model reasons: off, low, medium or high. The Shannon models set their own depth. Reasoning effort
• JSON object. With response_format: {"type": "json_object"} the answer is JSON without a prescribed shape. On /v1/chat/completions and /v1/responses. Structured Outputs
• JSON schema. With response_format of type json_schema the answer follows a schema you send. On /v1/chat/completions and /v1/responses. Structured Outputs
• Prompt caching. A prompt that starts like a recent one on the same model is read from cache, and the cached part is billed at 25% of the input price. It needs no request field. Prompt caching
• Web search. With web_search: true the API searches the web and gives the results to the model before it answers. On /v1/chat/completions and /v1/messages. It is a field of the request and has no flag in GET /v1/models. Built-in Web Search
• Streaming. With stream: true the reply arrives as server-sent events while the model writes. Every model streams on all three endpoints. Streaming

The same facts from GET /v1/models

GET https://api.shannon-ai.com/v1/models

GET /v1/models lists every model with a capabilities object, so a program can read these facts instead of this page. The endpoint needs no key.

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.shannon-ai.com/v1")

for model in client.models.list():
    print(model.id, getattr(model, "capabilities", None))

import OpenAI from "openai";

const client = new OpenAI({ apiKey: "YOUR_API_KEY", baseURL: "https://api.shannon-ai.com/v1" });

for await (const model of client.models.list()) {
  console.log(model.id, model.capabilities);
}

curl https://api.shannon-ai.com/v1/models

The id and capabilities of two entries of data:

[
  {
    "id": "shannon-3",
    "capabilities": {
      "tools": true,
      "vision": true,
      "reasoning": true,
      "reasoning_effort": false,
      "streaming": true,
      "json_schema": true
    }
  },
  {
    "id": "Kimi-K3-3BIT-REAP",
    "capabilities": {
      "tools": true,
      "vision": true,
      "reasoning": true,
      "reasoning_effort": true,
      "streaming": true,
      "response_format": true,
      "json_schema": true,
      "prompt_caching": true
    }
  }
]

• response_format and prompt_caching are flags of the hosted open-weight entries. A Shannon entry reports JSON output with json_schema alone, which stands for both JSON modes.
• vision is true for shannon-2-lite and shannon-2-pro: they read images in requests that also send tools or response_format.

Outside a model's row

A hosted open-weight model refuses an image or a JSON mode that is outside its row. The refusal comes before anything is drawn from your balance:

A Shannon model answers such a request with status 200: the capability outside its row takes no part in the answer. Error Handling