​

GET /v1/models

​

Authentication

• No API key — returns the full Openference model catalogue (all active, publicly listed models). Useful for clients that sync models during setup before credentials are configured.
• Valid API key — returns only the models your key and plan are allowed to use (respects model restrictions and your plan).
• Invalid or expired credentials — returns 401 (same as other API endpoints).

{
  "object": "list",
  "data": [
    {
      "id": "GLM-5.2",
      "object": "model",
      "created": 0,
      "owned_by": "openference",
      "permission": [],
      "root": "GLM-5.2",
      "parent": null,
      "context_length": 262144,
      "max_output_tokens": 131072,
      "pricing": {
        "prompt": "0.00000014",
        "completion": "0.00000028",
        "cache_read": "0.00000026"
      },
      "reasoning": {
        "supported": true,
        "control": "effort",
        "supported_efforts": ["high", "medium", "low"]
      },
      "quantization": "FP8"
    },
    {
      "id": "DeepSeek-V4-Pro-0813",
      "object": "model",
      "created": 0,
      "owned_by": "openference",
      "context_length": 1048576,
      "max_output_tokens": 384000,
      "reasoning": {
        "supported": true,
        "supported_efforts": ["max", "high"]
      }
    },
    {
      "id": "DeepSeek-V4-Flash-0731",
      "object": "model",
      "created": 0,
      "owned_by": "openference",
      "reasoning": {
        "supported": true,
        "supported_efforts": ["max", "high", "low"]
      }
    },
    {
      "id": "Kimi K2.7 Code",
      "object": "model",
      "created": 0,
      "owned_by": "openference",
      "reasoning": {
        "supported": true
      }
    }
  ]
}

​

Search

curl "https://api.openference.com/v1/models?search=glm-5.2" \
  -H "Authorization: Bearer $OPENFERENCE_API_KEY"

​

Context metadata

• context_length — the model’s total context window in tokens (OpenRouter-style field). Use it to configure your client’s context window setting (e.g. OpenCode limit.context, Cline “Context Window Size”) so the client doesn’t fall back to a small default and compact conversations early. Under dynamic context stacking, input and output share this total budget. For example, GLM-5.2 advertises context_length: 262144 and max_output_tokens: 131072. Shorter outputs allow the input prompt to scale dynamically up to the remaining context window.
• max_output_tokens — the maximum completion tokens the model supports per response.

​

Pricing

• pricing.prompt — the model’s per-token price for input (prompt) tokens, as a USD string. Multiply by the usage.prompt_tokens count returned by /v1/chat/completions to get the input cost for a request.
• pricing.completion — the model’s per-token price for output (completion) tokens, as a USD string. Multiply by usage.completion_tokens for the output cost.
• pricing.cache_read — when present, the per-token USD price for cached prompt tokens (prompt caching). Omitted for models that do not support prompt caching. When caching is supported but no separate cached rate is configured, this matches pricing.prompt.

​

Reasoning / thinking

​

Listing fields

• reasoning.supported — true when the model supports reasoning or extended thinking (including adaptive thinking on models that do not expose effort levels).
• reasoning.control — when present, how to control thinking for this model:
  • "effort" — send reasoning_effort (see supported_efforts)
  • "toggle" — on/off only (thinking.type, enable_thinking, or reasoning_effort: "none")
  • "adaptive" — MiniMax-style adaptive thinking toggle
  • "always_on" — reasoning always emitted; disable params are ignored
• reasoning.supported_efforts — when present, the effort levels you can request for that model (e.g. "high", "medium", "low", "max", "xhigh"). Omitted when the model uses always-on reasoning or on/off (adaptive) thinking only — those entries still include "supported": true.
• display_name — catalog display name when it differs from id.
• description — short catalog blurb when present.
• quantization — weight format from catalog (e.g. FP8, BF16).

​

Request parameters

​

Unified reasoning object (OpenRouter-compatible)

{ "reasoning": { "effort": "high" } }

• reasoning.effort — same ladder as reasoning_effort (reasoning_effort is an alias of reasoning.effort; sending both is fine, the explicit effort word wins).
• reasoning.max_tokens — a token budget for thinking (e.g. 8000), an alternative to an effort level. Mapped to the model’s native budget (budget_tokens / thinking_budget) with the model’s limits clamped in.
• reasoning.enabled — false disables thinking (when the model supports a toggle); true alone turns it on at a high default effort.
• reasoning.exclude — true asks the model to reason at full quality but omit the chain-of-thought from the response: no reasoning_content deltas on chat completions, no thinking blocks on Messages. The upstream still reasons and reasoning tokens are still billed as output. Not yet applied on the /v1/responses endpoint.

​

Responses

​

Flagship models (reference)