> four shots of a ceramic vase, warm side light
⏺ sequel - search_models(query: "cheap image model")
⎿ 3 models · flux-1-schnell $0.000634/run
⏺ sequel - get_model_schemas(model: "black-forest-labs/flux-1-schnell")
⎿ prompt, steps, width, height, seed
⏺ sequel - run_models(n: 4, steps: 4)
⎿ 4 images · $0.002536 · 4.1s




⏺ Four images, $0.002536, attributed to this session.
sequel mcp · image, video, audio · ? for shortcuts
Includes
Describe the job, not the model. Search matches on capability and sorts on price, so the cheapest one that can actually do it comes back first.
Phoenix 1.0$0.00583/tile_512 · ~9stext-to-image · text-rendering
Lucid Origin$0.006996/tile_512 · ~9stext-to-image · text-renderingNo model id, no provider, no parameter names. A sentence goes in and a priced run comes back — the same three calls your agent makes.
A composer types a job — An amber glass vase lit from above, black backdrop — and Sequel picks FLUX.1 [schnell] for it, at $0.000634 a run.




Grok Imagine Videovideo · $0.05/second
Grok Imagine Videovideo · $0.05/second
Lucid Originimage · $0.006996/tile_512
Lucid Originimage · $0.006996/tile_512
Phoenix 1.0image · $0.00583/tile_512
Grok TTSaudio · $0.000015/character
Phoenix 1.0image · $0.00583/tile_512
Grok TTSaudio · $0.000015/characterCalling a model is one line. The queue, the estimate, the delivery and the audit trail are the rest — and they run here, not in your codebase.
A video run comes back the moment it is accepted, with the model it chose, a task id and what the render will cost. Your agent keeps its turn; the work keeps going without it.
{ "status": "in_progress", "task_id": "task_9f2c7a1b", "model": "minimax/hailuo-2.3", "cost": { "usd": "0.280000", "estimated": true }, "eta_seconds": 80, "input": { "duration": 6, "resolution": "1080p", "fps": 24 }, "created_at": "2026-02-11T14:02:11.482Z" }
The estimate arrives with the task, not with the invoice. Show it, cap it, or refuse the run — all of it before the first second of compute.
A signed event lands on your endpoint when the asset is ready. Nothing polls, nothing sleeps, nothing holds a socket open to wait.
{ "type": "task.completed", "data": { "task": { "id": "task_9f2c7a1b", "status": "completed", "assets": [{ "url": "https://cdn.sequel.sh/task_9f2c7a1b/clip.mp4" }] } } }
A dropped connection costs you a notification, not the render. Redeploy, reconnect, come back tomorrow — the task is still there to fetch.
Who asked, from which client, on which key. A batch of clips arrives as a list you can audit, bill back, or quietly stop.