Models

Context is the most input tokens plus max_tokens one call may take; above it the call is refused with 400 input_too_large. Max output is the highest max_tokens; above it, 400 max_tokens_exceeded. Both models take the same fields. GET /models returns this table as JSON:

{
  "data": [
    {
      "id": "eleven-lite",
      "type": "model",
      "display_name": "Eleven Lite",
      "context_tokens": 32000,
      "max_output_tokens": 1024,
      "output_tokens_per_second": 120,
      "first_token_ms": 150,
      "pricing": {
        "input_usd_per_mtok": "2.00",
        "output_usd_per_mtok": "10.00"
      }
    },
    {
      "id": "eleven-max",
      "type": "model",
      "display_name": "Eleven Max",
      "context_tokens": 128000,
      "max_output_tokens": 4096,
      "output_tokens_per_second": 40,
      "first_token_ms": 600,
      "pricing": {
        "input_usd_per_mtok": "4.00",
        "output_usd_per_mtok": "20.00"
      }
    }
  ]
}

A fictional tokenizer counts 4 characters (Unicode code points) per token, rounded up. A call’s input is its system text plus every message, each counted on its own; the reply’s output tokens are its text at the same rate. The counts in usage are what you are charged.

So a 32,000-token context on Eleven Lite holds about 128,000 characters of prompt and reply together.

Deterministic: the same prompt and settings always give the same reply. Replies are placeholder text shaped by each model's profile and streamed at that model's speed. A fictional tokenizer counts input and output tokens at chars_per_token characters per token.

A reply’s first line says it is simulated. Its length is drawn from the model’s profile (150 to 650 tokens on Eleven Lite, 400 to 1,600 tokens on Eleven Max), its shape follows the profile, and it stops at the first stop sequence or at max_tokens. Replies never name a model or a vendor. Nothing you send is used to train anything, because nothing is trained.