LU Cloud API

Start

Create a key in the app under Account, Cloud API keys. It works on every cloud plan and every credit pack and spends the same credits at the same prices as the app. It cannot read or change your account. Send it as a bearer token:

Base URL   https://lu-labs.ai/api/inference/v1
Header     Authorization: Bearer lu_...

curl https://lu-labs.ai/api/inference/v1/chat/completions \
  -H "Authorization: Bearer $LU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "moonshotai/Kimi-K3.1-Pro", "messages": [{"role": "user", "content": "Hello"}]}'

from openai import OpenAI

client = OpenAI(base_url="https://lu-labs.ai/api/inference/v1", api_key="lu_...")

reply = client.chat.completions.create(
    model="moonshotai/Kimi-K3.1-Pro",
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)

image = client.images.generate(
    model="flux-schnell",
    prompt="a lighthouse at dawn, film photo",
    size="1024x1024",
)
print(image.data[0].url)

Endpoints

Chat

POST /chat/completions takes the OpenAI body and forwards the knobs it does not know. Every chat model of the app answers here, the direct models included. Each model keeps its own sampling range: a temperature or top_p beyond what the model handles well is brought into its range before the request goes out. An agent step that starts repeating itself is stopped, and the repeated part is not charged. Credits are reserved when a request starts and settled to the real use when it ends.

Examples: Kimi K3.1 Pro (moonshotai/Kimi-K3.1-Pro), DeepSeek V4.1 Flash 1017 (deepseek-ai/DeepSeek-V4.1-Flash-1017), GLM 5.3 (zai-org/GLM-5.3), Qwen 3.8 A95B (Qwen/Qwen3.8-2.4T-A95B). The full list with context window, tools, vision and credit rates is GET /models.

Images

POST /images/generations is the OpenAI image endpoint: model, prompt, n up to 4, size as WIDTHxHEIGHT, response_format url or b64_json. Our own fields: negative_prompt, seed, wait (seconds to wait for the result, default 50, 0 to return at once) and options for a studio image model. Done in time: 200 with data. Still rendering: 202 with an empty data and the jobs to poll. Every answer carries jobs.

Examples: Flux Schnell (fast) (flux-schnell), Flux Dev (quality) (flux-dev), Flux 2 Dev (flux-2-dev).

Every other Create job

POST /jobs takes the body the Create tab sends: { "kind", "model", "prompt", "params": { "op", ... } }. It runs the same checks, prices and credit reservation as the app and answers 202 with the job. Poll GET /jobs/{id} until status is succeeded, failed or canceled. A failed or canceled job gives its credits back. Files go up first through POST /files (multipart file and role: source, mask, train, audio or video) and are named by the returned path. For removebg, eraser and upscale the model may be left out.

Video examples: Wan 2.2 720p (wan-2.2-720p), Wan 2.2 Fast (wan-2.2-fast), LTX-2 (with audio) (ltx-2). A price that depends on the inputs (studio models, clip length, speech length) is best asked for with POST /jobs/quote; send the figure as params.max_credits and the job never books more. params.client_request_id (a UUID) makes a retry safe: the same id returns the job it already booked.

# 1. upload the start frame
curl https://lu-labs.ai/api/inference/v1/files -H "Authorization: Bearer $LU_API_KEY" \
  -F role=source -F file=@frame.png
# -> {"path": "<your-id>/<uuid>.png", ...}

# 2. price it, then start it with that price as the ceiling
curl https://lu-labs.ai/api/inference/v1/jobs/quote -H "Authorization: Bearer $LU_API_KEY" -H "Content-Type: application/json" \
  -d '{"kind": "video", "model": "wan-2.2-720p", "prompt": "the waves roll in",
       "params": {"op": "animate", "source_path": "<path>", "duration": 5}}'
# -> {"object": "job.quote", "credits": 1234, ...}

curl https://lu-labs.ai/api/inference/v1/jobs -H "Authorization: Bearer $LU_API_KEY" -H "Content-Type: application/json" \
  -d '{"kind": "video", "model": "wan-2.2-720p", "prompt": "the waves roll in",
       "params": {"op": "animate", "source_path": "<path>", "duration": 5, "max_credits": 1234}}'
# -> 202 {"object": "job", "id": "...", "status": "queued", "credits": 1234}

# 3. poll until status is succeeded, then fetch the file
curl https://lu-labs.ai/api/inference/v1/jobs/<id> -H "Authorization: Bearer $LU_API_KEY"
curl -L https://lu-labs.ai/api/inference/v1/jobs/<id>/content -H "Authorization: Bearer $LU_API_KEY" -o clip.mp4

Models

GET /models lists the chat models, the way OpenAI clients expect. GET /models?type=create lists every Create model with kind, ops, credits (base price, per duration, per second or per length), durations for video, mature, and for a studio model its inputs and options; tools prices removebg, eraser and upscale. ?type=all returns both in one list, each row with its type. A key only sees the models it may use.

Content setting per key

Every key has a content setting: Strict, Soft or Off, the same three as the account setting under Settings, Content policy. A new key follows the account. You can set a key stricter than the account, never wider: a key on Off counts only while the account itself is on Off, with the age confirmation that takes. Change it under Account, Cloud API keys.

• Strict: no model marked mature in Create and none of the direct chat models. A request for one is refused with 403 model_blocked_by_key_policy and nothing is charged.
• Soft: every model; Create prompts follow the soft rule of the account setting.
• Off: every model and no prompt rule beyond the hard bans.
• The hard bans hold on every setting and every key, in chat and in Create: material that abuses or exploits minors is never made, and a request for it is refused with 422 prompt_blocked.

Limits

• Chat: 60 requests per minute per account.
• Create jobs: 30 per minute per key, 60 per minute across all keys of an account. Each image of a run with n above 1 counts.
• Quotes: 60 per minute per key. Uploads: 60 per minute per account, 20 MB per file.
• Up to 5 active keys per account. A 429 carries Retry-After in seconds.
• Your plan's credits are the budget for all of it. Results stay for 7 days.

Errors

Every error is JSON with a plain English error sentence and, where useful, a code.