Update (2026-10-22): since the evening of 10/21, a key with no prepaid credit is no longer stopped for a small hold. When the estimate fits the free time left and only the 1.5x ceiling would not, the job runs free if that hold is 50 JPY or less, and the service covers any overrun. The 240-frame example below now runs with nothing held. Estimates also follow each key's own measured renders: on 10/22, another key that had rendered finals before got 488 GPU seconds ("about 2 minutes") for 240 frames and 3,661 GPU seconds ("about 15 minutes", 88 JPY held) for 1,800 frames.
If you let an AI agent render Blender scenes on a cloud GPU, the question you want answered first is not "how fast" but "what will this cost me, and can it run away". This is how JANCTION Render answers it, with estimates I took from the live service on 2026-10-21.
JANCTION Render is a cloud GPU render farm for Blender that AI agents call as an MCP server or HTTP API. Send a .blend file, a bpy script or a 3D file (glTF / FBX / USD) and get frames or an MP4 back. No local Blender or GPU is needed.
The order the agent follows
The MCP server's own instructions tell the agent to work in this order:
-
render_preview- a few frames in seconds, to check the scene - ask the user
-
render_estimate- when the job is long or the free time left is uncertain (no GPU used, nothing charged) -
render_final- withmax_yenas a ceiling if you want one
render_estimate takes the same arguments as the final render and returns time and money in one answer.
What the estimates look like
Same scene for all three (the public example ex12_logo_ring_spin.py), a new key with its welcome credit:
| request | GPU seconds | expected wait | free time covers | held from credit (ceiling) | charged if it runs as estimated |
|---|---|---|---|---|---|
| 48 frames, 1280x720 | 86 | "about 20 seconds" | all of it | 0 JPY | 0 JPY |
| 240 frames, 1920x1080 | 908 | "about 4 minutes" | all of it | 37 JPY | 0 JPY |
| 1,800 frames, 1920x1080 | 6,813 | "about 30 minutes" | 1,000 seconds | 872 JPY | 582 JPY |
A few things in that table matter for an agent:
- The hold is larger than the estimate. A render can take up to 1.5 times its estimate, so the service reserves that much from prepaid credit and only charges the time actually used. On the 240-frame job, the estimate fits the free time, but the 1.5x ceiling would not, so 37 JPY is held and 0 is expected to be charged.
- The answer includes the per-job and per-day limits that apply to the key (3,000 JPY per job and 10,000 JPY per day on this key), so the agent knows before it asks.
- When the scene has not been rendered as a final before, the estimate says so (
"confidence": "preliminary") and asks for a short test render before trusting it.
How close was it? Earlier the same day I rendered the 48-frame version of this scene for real: 89.39 GPU seconds against the 86 estimated. End to end it took 39 seconds, not 20: the estimate covers the queue and the render on the farm, and the 39 also includes the upload check and downloading 44 MB.
Ceilings and what happens without credit
-
max_yenonrender_finalrefuses the job before anything is reserved if its hold would exceed the ceiling.max_yen=0means "only if it fits the free time". - If the key has no prepaid credit for the part beyond the free time, the job is not started (except the small holds of 50 JPY or less described in the update above). The agent gets a top-up link to show the user; it cannot pay by itself.
- Prices are 0.1 JPY per GPU-second. A new key gets a welcome credit (100 JPY right away, 400 JPY more after saving a card that is not charged, valid 14 days); after that, previews are free up to 120 GPU-seconds a day per key.
Asking for an estimate yourself
Over HTTP it is one call:
curl -s https://render.janction.jp/v1/estimate \
-H "X-API-Key: $JANCTION_RENDER_API_KEY" -H "Content-Type: application/json" \
-d '{"scene_url": "https://render.janction.jp/samples/ex12_logo_ring_spin.py",
"kind": "final", "frame_start": 1, "frame_end": 240,
"width": 1920, "height": 1080, "output": "mp4"}'
The fields to read are human (the wait in words), seconds (GPU seconds), and in cost: covered_by_free_yen, max_charge_yen (the hold) and estimated_yen (the expected charge).
Limits worth knowing
- A final can be up to 3,000 frames in one job, rendered in parts of 240 under one job id.
- 1920x1080 on free time, up to 3840x2160 with prepaid credit.
- The farm has a small number of GPUs; when they are busy, the estimate's queue wait grows and the agent should tell the user rather than retry.
Connect it from Claude, ChatGPT or any MCP client at https://render.janction.jp/mcp. Current prices and limits: https://render.janction.jp/pricing and https://render.janction.jp/facts.json.
Top comments (0)