Venice Responses API (Alpha)

POST /api/v1/responses is Venice's OpenAI-compatible Responses endpoint. It returns a structured, typed output array instead of a single message.content string — ideal for agents that need to separate reasoning, messages, tool calls, and built-in tool events.

Alpha. Access is gated behind the responsesApiEnabled flag on Bearer API keys (staff-only during beta). x402 wallet auth bypasses this flag — you can pay per request without the flag. Schemas may change.

Use when

• You need the OpenAI Responses-style response shape (output[] with typed type: "reasoning" | "message" | "function_call" | "web_search_call" blocks) for a client library that expects it.
• You want clean separation of reasoning vs message vs tool-call output.
• You want streaming via SSE with typed events.

Otherwise use venice-chat — it has more features, more models, and full Venice parameters.

Limitations vs /chat/completions

Limitation	Detail
Stateless	No conversation persistence across requests. Send the full history each call.
E2EE models default to rejection	E2EE-capable models return 400 unless you pass venice_parameters.enable_e2ee: false (TEE-only mode). For end-to-end encrypted inference with E2EE headers, use /chat/completions.
Subset of venice_parameters	character_slug, enable_e2ee, enable_web_search, enable_web_scraping, enable_web_citations, include_venice_system_prompt, include_search_results_in_stream are supported. strip_thinking_response, disable_thinking, enable_x_search are not wired through in Alpha.
Access gated by feature flag	Bearer keys without responsesApiEnabled get 401. x402 requests are allowed (pay-per-call).

Authentication

Same as the rest of the API — either Authorization: Bearer <key> or SIGN-IN-WITH-X: <SIWX>. See venice-auth.

Minimal request

curl https://api.venice.ai/api/v1/responses \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org-glm-5-1",
    "input": "Explain why the sky is blue in one paragraph."
  }'

input accepts:

• a plain string, or
• an array of typed input items (similar to chat/completions message parts) for multi-turn or multimodal history.

Response shape

{
  "id": "resp_abc123",
  "object": "response",
  "created_at": 1735689600,
  "model": "zai-org-glm-5-1",
  "status": "completed",
  "output": [
    {
      "type": "reasoning",
      "id": "rs_1",
      "summary": ["I considered Rayleigh scattering..."],
      "encrypted_content": "..."
    },
    {
      "type": "message",
      "id": "msg_1",
      "status": "completed",
      "role": "assistant",
      "content": [{
        "type": "output_text",
        "text": "The sky is blue because...",
        "annotations": [{
          "type": "url_citation",
          "url": "https://example.com/rayleigh",
          "title": "Rayleigh scattering",
          "start_index": 42,
          "end_index": 99
        }]
      }]
    },
    {
      "type": "function_call",
      "id": "fc_1",
      "call_id": "call_abc",
      "name": "get_weather",
      "arguments": "{\"city\":\"Paris\"}",
      "status": "completed"
    },
    {
      "type": "web_search_call",
      "id": "ws_1",
      "status": "completed"
    }
  ],
  "usage": {
    "input_tokens": 20,
    "input_tokens_details": {"cached_tokens": 0},
    "output_tokens": 80,
    "output_tokens_details": {"reasoning_tokens": 40},
    "total_tokens": 100
  }
}

Top-level status ∈ completed | failed | in_progress | cancelled. On failed, error.code and error.message are populated.

Output block types

type	Purpose
reasoning	Thought process from reasoning models. summary[] holds human-readable text; encrypted_content holds opaque signatures — round-trip verbatim for multi-turn tool calls.
message	Main text output. content[].type === "output_text", plus annotations[] for url_citation entries from web search.
function_call	Tool call: name, stringified-JSON arguments, call_id.
web_search_call	Sentinel showing the built-in web_search tool fired; use alongside url_citation annotations on messages.

Match tool outputs back by call_id when continuing the turn.

Common request fields

Field	Notes
model	Required. Model ID, trait, or compatibility mapping. Feature suffixes allowed (see venice-chat).
input	Required. String or input-items array. To set system/developer context, include a leading message with role: "system"/"developer" in the input array.
tools	Array of {type:"function",function:{...}} or built-in {type:"web_search"} — availability depends on the model.
tool_choice	"auto" / "required" / "none" / {type:"function",function:{"name":"..."}}.
reasoning.effort	Reasoning effort hint for thinking models ("low" | "medium" | "high"). Mapped to reasoning_effort.
temperature, top_p, max_output_tokens	Standard generation controls. max_output_tokens maps to max_tokens.
web_search	Boolean shortcut for enabling web search, equivalent to adding {"type":"web_search"} to tools or setting venice_parameters.enable_web_search.
include	Array of additional response fields to include (OpenAI compat).
fallbacks	Up to 10 entries. Anthropic beta parameter for Claude Fable 5 server-side refusal fallback; forwarded only on direct Anthropic routes.
stream	Boolean. SSE response with typed events (response.created, response.output_item.added, response.output_text.delta, response.completed, …).
venice_parameters	Subset listed above. Example: {"character_slug":"alan-watts","enable_web_search":"on"}.

Silently dropped generation controls. n, stop, seed, and prompt_cache_key are not in the Alpha schema and are not translated to /chat/completions. The request body is permissive, so sending them does not error — they just never reach inference. If you need reproducible sampling (seed), stop sequences, or explicit cache routing, stay on /chat/completions.

Fields commonly found in OpenAI's Responses API that are not in Venice's Alpha schema (and silently ignored or rejected by Zod): instructions, metadata, parallel_tool_calls, response_format, store, previous_response_id, background. For response_format / JSON-schema structured output, use /chat/completions.

Streaming

With stream: true, the response is an SSE stream of typed events. Typical flow:

event: response.created
event: response.output_item.added        # type=reasoning
event: response.reasoning.delta
event: response.output_item.added        # type=message
event: response.content_part.added
event: response.output_text.delta
event: response.output_text.delta
event: response.output_item.done
event: response.completed

Consume events in order and reconstruct output[] client-side; the shape on response.completed matches the non-streamed response exactly.

Authentication & error responses

• 400 — bad request; also returned when an E2EE-capable model is used without venice_parameters.enable_e2ee: false.
• 401 — auth failed, or Bearer key lacks responsesApiEnabled, or the model is Pro-only and you're on an INFERENCE key / x402 wallet.
• 402 — insufficient balance. Bearer → { error: "INSUFFICIENT_BALANCE" }. x402 → PAYMENT_REQUIRED with topUpInstructions and siwxChallenge (see venice-x402).
• 429 — rate-limited.
• 500 — inference failed.

X-Balance-Remaining is on 200 responses when using x402 auth; PAYMENT-REQUIRED header on 402.

Migration notes

• Port messages → pass as input (string, or typed array with leading {role:"system"|"developer", content:"..."}).
• venice_parameters.character_slug → supported; pass inside venice_parameters or as a model feature suffix (:character_slug=alan-watts).
• venice_parameters.enable_web_search → pass inside venice_parameters, or append :enable_web_search=on to the model ID, or add {"type":"web_search"} to tools.
• venice_parameters.strip_thinking_response / disable_thinking → not supported on /responses in Alpha; stay on /chat/completions for these.
• Full E2EE flow (E2EE request headers + encrypted response) → stay on /chat/completions. For TEE-only inference on an E2EE-capable model, pass venice_parameters.enable_e2ee: false here.
• response_format / JSON-schema structured output → stay on /chat/completions.