GET /v1/models

Lists the enabled models in the OpenAI list format, with prices and limits added. It needs no API key. The same data is published at /models.json.

Placeholders: $API_BASE is https://api.abliteratedllm.dev/v1, $API_ORIGIN is https://api.abliteratedllm.dev, and $YOUR_KEY is your API key.

curl $API_BASE/models

Abridged response (the table below is live):

{
  "object": "list",
  "data": [
    {
      "id": "glm-5.3-flash-abliterated",
      "object": "model",
      "created": 1793664000,
      "owned_by": "…",
      "name": "GLM-5.3 Flash Abliterated",
      "context_length": 262144,
      "max_output_tokens": 65536,
      "pricing": { "input": "0.0977", "output": "0.3421", "cached_input": "0.0098" },
      "provider_label": null,
      "precision": "FP8"
    }
  ]
}

Fields

Prices are strings so that no precision is lost: parse them as decimals, not floats.

Current models

Precision

glm-5.3-flash-abliterated is served in native FP8 (block-FP8, e4m3): the precision Z.ai releases GLM-5.3 Flash in. Zero additional quantization: it is never compressed to INT8, INT4 or FP4.

Official DeepSeek API models

Models whose provider_label is "Official DeepSeek API" are DeepSeek's stock models. Requests go through our gateway, with the same key, balance, limits and billing, and are then answered by DeepSeek's official API. They differ from our abliterated models:

• Not abliterated. They keep DeepSeek's own refusals. We add no system prompt and don't rewrite your messages.
• DeepSeek may decline requests. DeepSeek can refuse some requests on its side under its Terms of Use. Our own safety check is the same as on every other model.
• Data handling. Prompts and outputs for these models are sent to DeepSeek, processed and stored in the People's Republic of China under the DeepSeek Privacy Policy, and may be used by DeepSeek to train its models. Our no-storage promise covers our own systems only. Details are in the privacy summary.
• Versions move. DeepSeek can switch the model behind each of its API ids to a newer version without notice.
• Thinking is on by default. Turn it off with reasoning_effort: "none" on chat completions or thinking: {"type": "disabled"} on messages. While the model thinks, DeepSeek ignores sampling settings such as temperature. response_format supports only text and json_object.

How a request is billed

Balances and costs are kept in millionths of a dollar ($0.000001). For each request:

1. Hold. Before forwarding, we reserve the most the request could cost: the estimated input (request size in bytes divided by 3, in tokens) at the input price, plus the full output budget at the output price. The hold is subtracted from your available balance while the request runs. Rate limits and trial explains what happens when the hold doesn't fit.
2. Settle. When the response finishes, the hold is released and you pay the usage the model reported: input, cached input and output tokens, each at its price per 1M, each rounded up to the next $0.000001. The formulas per endpoint are on the chat completions, messages and responses pages.
3. Exceptions.
   • If you disconnect before the response ends, streaming or not, generation stops and you pay an estimate. Input is the hold's input estimate, capped at the model's context length. Output is the output already produced: its size in UTF-8 bytes divided by 3, at least one token per streamed piece, capped at the output budget.
   • If a response ends without a usage report, the same estimate applies.
   • Requests that fail before producing output (validation errors, rate limits, capacity_exhausted, upstream_error) cost nothing.

Your per-request history (time, key, model, token counts, cost, status and latency) is in the dashboard for 30 days, then rolled up into daily totals.