Sync routes

One HTTP round-trip after payment. Paths mirror Venice under /api/v1/*.

Request bodies follow Venice / OpenAPI schemas. model must be an exact id from GET /api/v1/models (or resolved via /models/traits). Prefer the cheapest suitable model unless the user asks otherwise.

200 response shapes are recorded from live probes (outputFixtures) and stamped onto OpenAPI responses.200 / Bazaar info.output. Do not invent field names. Observed JSON (not exhaustive):

Edit / upscale / background-remove / audio/speech → raw binary (image/png or audio/*).

Reasoning / thinking models

Venice thinking models (e.g. GLM, Kimi including kimi-k3-1-pro which reasons at max effort by default, Claude Opus, GPT-5.4 Pro variants) may return more than choices[0].message.content:

Do not treat empty content alone as failure. If reasoning_content or reasoning_details is present, the completion succeeded. With a small max_tokens / max_completion_tokens, thinking can consume the whole budget (finish_reason: "length") and leave content empty.

Controls (request venice_parameters):

Example — smoke / ping that always wants visible text:

{
  "model": "zai-org-glm-5-2",
  "messages": [{ "role": "user", "content": "ping" }],
  "max_tokens": 128,
  "venice_parameters": { "disable_thinking": true }
}

For long music/video/Voice Changer jobs, use the async lifecycle instead of sync TTS.

Notable request fields

• Responses input arrays may include { "type": "additional_tools" } items to add tools mid-conversation. Optional top-level instructions, tools (function / custom with namespace / tool_search), tool_choice by name, and reasoning.summary. Image parts accept image_url.detail auto|low|high|original.
• Image generate: quantity is variants (1–4), not OpenAI n. Charge = list unit × variants. You pay for every image requested. If Venice omits some variants for an explicit provider content-policy reject, the 200 still settles and x-venice-content-violation-count (forwarded) is the omitted count. A full-prompt 422 ImageContentViolationError (retryable, optional suggested_prompt / recommended_model) is returned unpaid (settled: false). Also: enhance_prompt, style_references (only models with supportsStyleReferences). safe_mode defaults true (turning it off does not require a Venice Pro subscription). The x-venice-is-blurred response header reports whether the returned image was blurred. Optional transparent_background (native generate/edit/multi-edit) returns an alpha channel — only models with model_spec.constraints.supportsTransparentBackground=true, and format / output_format must be png or webp. OpenAI-shaped /images/generations uses n (Venice max 1) and background: transparent | opaque | auto instead. Same unpaid 422 refusal.
• Image edit: enhance_prompt. Same safe_mode / x-venice-is-blurred as generate. Same optional transparent_background (and output_format png/webp) as generate.
• TTS: response_format must be in that model’s model_spec.supported_formats.
• Video queue: prompt max is model-specific (up to 32000). Resolutions include 768p when that model’s model_spec.constraints.resolutions lists it. Swagger now lists aspect 2:1 (still only valid when that model’s model_spec.constraints.aspect_ratios includes it). Seedance 2.x reference-to-video: duration -1/auto and aspect adaptive/auto match the source; send consents when the model returns needs_consent. MiniMax H3 Max Recast (minimax-h3-max-recast) lists duration Auto only — auto / -1 are accepted. bitrate_mode is not in current Venice swagger — omit it.