Venice Chat Completions

POST /api/v1/chat/completions is Venice's main text endpoint. It's OpenAI-compatible, plus a venice_parameters object for Venice-only features. Auth is a Bearer API key or an x402 wallet (SIGN-IN-WITH-X); see venice-auth.

Use when

• You need LLM text generation, with or without tools, with or without streaming.
• You want multimodal inputs (images, audio, video, documents) to a capable model.
• You want Venice-specific features: web search, web scraping, citations, E2EE, characters, xAI X/Twitter search, thinking control.
• You need prompt caching for large system prompts or long documents.
• You need structured (json_schema) output.

For the OpenAI Responses-style shape (typed output[] blocks), see venice-responses. To pick a model, see venice-text-routing.

Minimal request

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org-glm-5-1",
    "messages": [{"role": "user", "content": "Why is the sky blue?"}]
  }'

OpenAI SDK (Python) — Venice-only fields go in extra_body:

import os
from openai import OpenAI
client = OpenAI(api_key=os.environ["VENICE_API_KEY"], base_url="https://api.venice.ai/api/v1")
resp = client.chat.completions.create(
    model="zai-org-glm-5-1",
    messages=[{"role": "user", "content": "Summarize today's AI news."}],
    extra_body={"venice_parameters": {"enable_web_search": "auto", "include_venice_system_prompt": False}},
)

Response shape is the standard OpenAI chat.completion object (id, object: "chat.completion", created, model, choices[].message, choices[].finish_reason, usage) plus:

• cost: {usd, diem} — the request's cost split by the currency it was charged in (bundled credits count as USD). 0 when the response has no output tokens (those responses are not billed). Omitted if the cost can't be computed, and currently also when the request was charged to earned credits.
• venice_parameters — the effective Venice settings, plus web_search_citations[] (an empty array when no search ran).
• usage.prompt_tokens_details.{cached_tokens, cache_creation_input_tokens} and usage.completion_tokens_details.reasoning_tokens when the provider reports them.

choices[].finish_reason is one of stop, length, tool_calls, or content_filter. content_filter is a 200 in which the model refused or the output was cut (e.g. a Claude refusal); it is not a 422, so check for it before trusting message.content. Some providers block mid-stream instead; the stream then ends with an in-band error chunk (type: "content_filter_error", code: "content_blocked_by_provider") followed by [DONE].

system_fingerprint is always stripped. With stream: true, responses come as SSE data: lines in chat.completion.chunk format.

The request body

The top-level schema is strict: unknown top-level fields return 400. (A few compatibility aliases are rewritten before validation: input → messages, max_output_tokens → max_tokens, web_search: true|false → enable_web_search, promptCacheKey → prompt_cache_key.)

Core fields (OpenAI-compatible)

Field	Notes
model	string — model ID, trait (e.g. default, default_code), or compatibility mapping. Required. Feature suffixes allowed (see below). Lookups also tolerate dots/underscores/spaces (kimi k2.6 → kimi-k2-6) and a venice- prefix (for IDEs that hijack claude-* names).
messages	array of system / developer / user / assistant / tool messages. Required, min 1. Assistant messages with neither content nor tool_calls are silently dropped.
temperature (0–2), top_p (0–1), top_k (int ≥ 0), min_p (0–1), min_temp, max_temp (0–2)	sampling controls. Some models publish defaults in model_spec.constraints
repetition_penalty (≥ 0), frequency_penalty, presence_penalty (−2..2)	repetition controls
max_completion_tokens / max_tokens (deprecated)	integers. Output cap including reasoning tokens. Above the model's model_spec.maxCompletionTokens → 400 on models with an enforced API cap. max_tokens ≤ 0 is ignored; max_tokens is ignored when max_completion_tokens is set
n	number of choices (default 1; you pay for all choices)
seed	positive integer
stop / stop_token_ids	string or 1–4 strings / array of token IDs
stream, stream_options.include_usage	SSE streaming (Venice sends the usage chunk even without include_usage — see Streaming)
response_format	{type:"json_schema", json_schema:{name, schema, strict}} (preferred), {type:"json_object"}, or {type:"text"}
tools, tool_choice, parallel_tool_calls	function calling
logprobs, top_logprobs (int ≥ 0)	log-probabilities
reasoning_effort / reasoning.effort	none | minimal | low | medium | high | xhigh | max. reasoning_effort wins if both are set
reasoning.enabled	false disables reasoning on supported models. Ignored when an effort is set
reasoning.summary	auto | concise | detailed
prompt_cache_key	cache-routing hint. If omitted, Venice derives a stable key per API user
prompt_cache_retention	default | extended | 24h. extended and 24h extend retention to 24 hours on supported models
verbosity / text.verbosity	low | medium | high | auto (both placements accepted). Non-reasoning OpenAI models accept only medium / auto (400 otherwise)
anon_user_id	optional end-user identifier (see below)
fallbacks	up to 10 {model} entries. Anthropic beta parameter for Claude Fable 5 server-side refusal fallback. Forwarded only on direct Anthropic routes, ignored elsewhere
include, metadata	accepted for OpenAI compatibility, removed before validation and not forwarded
user, store	accepted and discarded (OpenAI compat). user is not an alias of anon_user_id

anon_user_id — identifies your end user; Venice combines it with your Venice user id when attributing the request upstream. Trimmed (a blank value is treated as absent); 1–128 characters; printable ASCII only (0x20–0x7E); must not contain ||. Violations → 400. Also accepted on /responses.

Capability gates (400 before inference)

You send	Model must have (model_spec.capabilities)
image_url parts	supportsVision
input_audio parts	supportsAudioInput
video_url parts	supportsVideoInput
tools, tool_choice, parallel_tool_calls: true	supportsFunctionCalling
response_format other than text	supportsResponseSchema
logprobs: true or any top_logprobs	supportsLogProbs

Each rejection is a 400 whose issues[] entry for that field gives the reason; read issues, not just the top-level error.

OpenAI reasoning models (e.g. openai-gpt-52) also reject seed, stop, n ≠ 1, and non-zero presence_penalty / frequency_penalty with 400 (the issues[] entry for that field reads "<field> is not supported by this model").

venice_parameters (Venice-only)

All optional. Unknown keys inside venice_parameters are dropped.

Field	Type	Default	Effect
character_slug	string	—	Apply a published Venice character (the "Public ID" on its page). Unknown slug → 404. See venice-characters.
strip_thinking_response	bool	false	Strip reasoning from the response (<think> blocks and reasoning_content) on reasoning models.
disable_thinking	bool	false	Disable thinking on supported reasoning models and strip reasoning. On models that can't turn reasoning off, reasoning still runs (some drop to their lowest effort) but is stripped from the response; those reasoning tokens may still be billed.
enable_e2ee	bool	true	On E2EE-capable models, use E2EE when E2EE headers are present. false forces TEE-only mode.
enable_web_search	"off" / "auto" / "on"	"off"	Venice web search. on always searches; auto lets a classifier decide.
enable_web_scraping	bool	false	Scrape URLs found in the latest user message. When URLs are found, scraping replaces web search for that request.
enable_web_citations	bool	false	Ask the model to cite sources as ^1^ / ^1,3^.
include_search_results_in_stream	bool	false	Experimental. Streaming only: emit a choices: [] chunk carrying venice_parameters.web_search_citations at the end of the stream, before data: [DONE].
return_search_results_as_documents	bool	—	Also surface search results as a synthetic tool call (see Web search).
include_venice_system_prompt	bool	true	Prepend Venice's system prompt to yours. Set false for full control.
enable_x_search	bool	false	xAI native web + X search on models with supportsXSearch (Grok). Ignored on other models. Billed per search (~$0.01).

Model feature suffixes

Some venice_parameters can be set on the model string — useful when the client (OpenAI SDK, LangChain, an IDE) can't send venice_parameters:

<model-id>:<key>=<value>[&<key>=<value>…]

Values are URL-decoded; booleans are true / false. Suffixes override the same keys in venice_parameters. Supported keys (exact match):

Key	Values	Maps to
enable_web_search	on / off / auto	venice_parameters.enable_web_search
enable_web_citations	true / false	venice_parameters.enable_web_citations
enable_web_scraping	true / false	venice_parameters.enable_web_scraping
include_venice_system_prompt	true / false	venice_parameters.include_venice_system_prompt
include_search_results_in_stream	true / false	venice_parameters.include_search_results_in_stream
return_search_results_as_documents	true / false	venice_parameters.return_search_results_as_documents
character_slug	string	venice_parameters.character_slug
strip_thinking_response	true / false	venice_parameters.strip_thinking_response
disable_thinking	true / false	venice_parameters.disable_thinking

Unknown keys are silently ignored. enable_e2ee and enable_x_search are not suffix keys. Suffixes only apply on /chat/completions — /responses resolves the model but ignores the suffix.

zai-org-glm-5-1:enable_web_search=on
kimi-k2-6:strip_thinking_response=true&enable_web_search=auto
zai-org-glm-5-1:character_slug=alan-watts

Messages and modalities

messages[].content is a string or an array of typed parts. Roles: user, assistant, tool, system, developer. Only user messages may carry image / audio / video / file parts; system / developer / assistant array content is text-only. Anthropic-style tool_use / tool_result blocks and Cursor-style {type:"image"} parts are converted automatically.

Per-message caps: 10 image_url, 5 input_audio, 3 video_url, 5 file parts. Per-request cap: 3 video_url parts total.

Images (image_url)

{
  "model": "kimi-k2-6",
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "What's in this image?"},
      {"type": "image_url", "image_url": {"url": "https://example.com/cat.jpg"}}
    ]
  }]
}

• url is a public http(s) URL or a data:image/...;base64,... URL (the example.com placeholder above fails validation — swap in a real image). Remote URLs are fetched once for validation: redirects are refused, the body must be ≤ 25 MB, the Content-Type must be PNG, JPEG, WebP, HEIF/HEIC, or AVIF, and the image must decode and be ≥ 64 px on each side (data URLs get the same decode + size check). Failures → 400 "Supplied image did not pass validation checks."
• Models with supportsMultipleImages: true keep images across the whole conversation (maxImages advertises the model's per-request limit; Venice itself enforces the 10-per-message cap). Single-image vision models keep images only from the last image-bearing message; earlier images are removed.

Audio (input_audio)

{"type": "input_audio", "input_audio": {"data": "<base64>", "format": "wav"}}

format: wav (default), mp3, aiff, aac, ogg, flac, m4a, pcm16, pcm24. Audio must be inline base64 — URLs are not supported.

Video (video_url)

{"type": "video_url", "video_url": {"url": "https://www.youtube.com/watch?v=..."}}

• YouTube watch/share/embed links are accepted without a pre-fetch (provider support varies).
• data:video/...;base64,... URLs must declare video/mp4, video/mpeg, video/quicktime, video/mov, or video/webm.
• Other remote URLs must return 2xx without redirects and one of those Content-Types. Failures → 400 "Supplied video did not pass validation checks."
• More than 3 video_url parts in a request → 400 (checked before any URL is fetched).

Documents (file)

{"type": "file", "file": {"file_data": "data:application/pdf;base64,JVBERi0...", "filename": "report.pdf"}}

file_data is a data URL or a public URL. PDF, EPUB, DOCX, PPTX, XLSX, XLS, plain text, Markdown, CSV, JSON, and most source-code files are extracted to text server-side (so any text model works); image files become image_url parts. A file that fails extraction becomes an inline [Error processing file …] text part rather than a request error. Not allowed on E2EE requests.

Prompt caching (cache_control)

Any content part can carry {"cache_control": {"type": "ephemeral"}} or {"type": "ephemeral", "ttl": "1h"}. Explicit markers matter for models that require them (Claude); other models cache automatically on prefix matches. Those models allow at most 4 breakpoints per request; Venice adds its own to the system prompt and conversation history only while your markers leave slots free. Pair with a stable prompt_cache_key for consistent routing. Cache read / write prices are per model (model_spec.pricing.cache_input / cache_write); hits are reported in usage.prompt_tokens_details.

Tools & function calling

{
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Get current weather for a city",
      "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]},
      "strict": true
    }
  }],
  "tool_choice": "auto"
}

• tool_choice: "auto", "required", "none", or {"type":"function","function":{"name":"get_weather"}}. {"type":"auto" | "none" | "required"} is normalized to the string form.
• parallel_tool_calls defaults to true — be ready to run several calls before replying.
• Reply with one {"role":"tool","tool_call_id":"...","content":"..."} message per call, then call again.
• Flat ({type, name, parameters}) and Anthropic ({name, input_schema}) tool definitions are converted to the nested format.
• When the message carries tool calls, Venice reports finish_reason: "tool_calls" rather than stop (a length cut-off stays length).
• The schema also accepts {"type":"web_search"} / {"type":"x_search"} tool entries, but they do not turn on Venice search or xAI X search. Use venice_parameters.enable_web_search / enable_x_search instead.

Reasoning models

{
  "model": "zai-org-glm-5-1",
  "reasoning": {"effort": "medium"},
  "messages": [{"role": "user", "content": "Prove there are infinitely many primes."}]
}

• Reasoning arrives in message.reasoning_content (delta.reasoning_content when streaming); <think> tags are removed from content. Some providers return encrypted or summarized reasoning.
• Supported effort values are per model: check model_spec.capabilities.supportsReasoningEffort, reasoningEffortOptions, and defaultReasoningEffort. On most Claude models and on OpenAI GPT models, a value outside reasoningEffortOptions → 400 (none is always accepted as a Venice-level off switch; on OpenAI models minimal is also accepted and maps to the lowest supported level). Other models are not pre-validated, so stick to reasoningEffortOptions.
• To turn thinking off, prefer reasoning: {"enabled": false} or venice_parameters.disable_thinking: true — both degrade gracefully on mandatory-reasoning models. reasoning_effort: "none" also disables thinking where the model allows it, but some mandatory-reasoning models reject it; prefer the two switches above.
• Some models return reasoning_details[] (and Gemini via native transport returns thought_signature) on the assistant message. Pass them back verbatim on the next turn, especially in tool loops, to preserve thought signatures.
• Reasoning tokens count toward max_completion_tokens and are billed as output.

Structured output (response_format)

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "person",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {"name": {"type": "string"}, "age": {"type": "number"}},
        "required": ["name", "age"],
        "additionalProperties": false
      }
    }
  }
}

Put the JSON Schema under json_schema.schema (OpenAI shape) — Claude, Gemini, and Grok models only read it from there. json_object is deprecated and ignored by Claude models. Requires supportsResponseSchema.

E2EE (end-to-end encryption)

For models with supportsE2EE: true (the e2ee-* IDs), following the TEE & E2EE guide:

1. GET /api/v1/tee/attestation?model=<id>&nonce=<64 hex chars> (no auth needed, 10 req/min/IP). Verify it and take the model's public key.
2. Generate a per-session secp256k1 key pair. Encrypt every user and system message with ECDH → HKDF-SHA256 → AES-256-GCM.
3. Send the request with X-Venice-TEE-Client-Pub-Key and X-Venice-TEE-Model-Pub-Key (secp256k1 hex keys), X-Venice-TEE-Signing-Algo: ecdsa, and stream: true (the guide requires streaming). Malformed headers → 400 with {"error":{"message":"Invalid E2EE headers: …","type":"invalid_request_error"}}.
4. Decrypt the streamed content with your private key.

On E2EE requests Venice injects nothing: no Venice system prompt, character, web search, or scraping. file parts → 400. The guide also lists function calling as unsupported. Without E2EE headers (or with enable_e2ee: false) the same model runs in TEE-only mode. TEE responses carry X-Venice-TEE: true and X-Venice-TEE-Provider. E2EE is not available on /responses.

Streaming

{"stream": true}

• text/event-stream, one data: {chat.completion.chunk} per event, terminated by data: [DONE].
• Venice requests usage from the model and emits it in a choices: [] chunk with usage (and cost) before [DONE], whatever stream_options.include_usage says. Usage is not repeated on content chunks.
• If the upstream fails after headers are sent, you get an in-band data: {"error": {...}} chunk (e.g. code: "model_overloaded" with retry_after, upstream_error, or content_blocked_by_provider) followed by [DONE].
• Web-search citations are not in the stream unless include_search_results_in_stream: true, in which case a choices: [] chunk with venice_parameters.web_search_citations arrives at the end of the stream, before [DONE].
• With return_search_results_as_documents: true, the synthetic web_search_call tool call (see Web search) is streamed as a delta.tool_calls chunk before the content.

Web search

• Non-streaming responses include venice_parameters.web_search_citations[] with url, title, content (snippet), and date. Add enable_web_citations: true to get ^n^ markers in the text.
• Search is skipped on E2EE requests, when the last user message contains an image, and when scraping found URLs in that message.
• return_search_results_as_documents: true adds a synthetic tool call {id:"web_search_call", type:"function", function:{name:"web_search", arguments:"{\"documents\":[{id,title,url,snippet,published_at}]}"}} and sets finish_reason: "tool_calls". Don't try to execute it.
• Billing: $0.01 per search-augmented request, $0.01 per scraped URL, $0.01 per xAI X search.

Error handling specifics

Status	When
400	Invalid body (error: "Invalid request parameters", per-field reasons in issues[]), capability rejections, invalid image/video, too many parts, context length exceeded, token cap exceeded, bad E2EE headers
401	Invalid API key or SIWX sign-in; also a model that requires a paid subscription
402	No credentials at all (x402 discovery body — not 401), insufficient balance, or API-key spend limit. x402 insufficient balance: code: "PAYMENT_REQUIRED" body with topUpInstructions + siwxChallenge, and a PAYMENT-REQUIRED header (venice-x402)
403	Model not allowed by the API key's modelPrivacy, region-restricted model, or provider restriction
404	Unknown model (often with a "Did you mean" hint) or character_slug
413	Payload too large
422	Content-policy violation (Venice or provider)
429	Rate limit exceeded ("Rate limit exceeded" or the error-budget message), or model overloaded ("The model is currently overloaded…" with a Retry-After header). Both carry x-ratelimit-* headers, so tell them apart by Retry-After and the message
500 / 503 / 504	Inference failed / model offline / upstream timeout

See venice-errors for shapes and retry strategy.

Common gotchas

• max_tokens is deprecated — use max_completion_tokens, and stay under model_spec.maxCompletionTokens.
• Unknown top-level fields are rejected (400); unknown venice_parameters keys are silently dropped.
• Image and video URLs must be publicly reachable without redirects. Signed S3 URLs that redirect or localhost URLs fail.
• Audio cannot be a URL — always base64.
• Single-image vision models drop older images each turn; put images in the last user message.
• Round-trip reasoning_details / thought_signature unchanged in multi-turn tool loops.
• character_slug adds the character's system prompt ahead of yours. Venice's own prompt is still included unless the character uses a custom system prompt or you set include_venice_system_prompt: false.
• web_search / x_search tool entries are not Venice search — use venice_parameters.
• OpenAI reasoning models reject seed, stop, n > 1, and non-zero penalties.