Input
Original
The Symbioza mark: a 340-pixel image, the only input.
symbioza-mark-340.webpDemo / Logo → 3D
One logo.
One ChatGPT conversation.
The agent handed the heavy 3D work to Symbioza as GPU jobs and got real meshes back.
- 01Ask in your agentyou + agent
- 02Review the estimatefree
- 03Approve a spending limityou
- 04Symbioza runs the GPU workSymbioza
- 05Get the files backmeshes + previews
01 / From a logo to actual geometry
Original. TripoSR.
TRELLIS. Refined.
The two model stages ran as GPU jobs through Symbioza. The final polish was done by hand in Blender, locally.
GPU job · Symbioza
TripoSR
A first reconstruction. You can recognise it, but the cap is rough and the roots break apart.
45 files back
GPU job · Symbioza
TRELLIS
A better model, run three times with different seeds. Seed 42 was kept for its cap and stem.
3 candidates · 22 files back
Local · not GPU compute
Refined
The seed-42 mesh, reshaped in Blender: brain-like folds on the cap, new roots, materials.
finished by hand02 / The ask
Ask for it
the way you'd ask a person.
Your agent doesn't shop for a GPU. It asks Symbioza for a price, waits for your yes, then brings the files back into the conversation.
ESTIMATE FIRST
03 / The receipts
Every GPU attempt, billed.
Four attempts. Two delivered. The two first tries were billed too, and they are listed here.
GPU JOBS · 2026-10-032 delivered
| Execution | Stage | Charge |
|---|---|---|
| exec-67dd4994 | TripoSRfirst try; failed in dependency install | $0.02417final |
| exec-7eed4e9a | TripoSRretry; delivered 45 of 45 files | $0.05783602provisional |
| exec-79c4bd05 | TRELLISfirst try; not ready in 480 s, never ran | $0.001final |
| exec-ece6ba27 | TRELLISretry, three seeds; delivered 22 of 22 files | $0.09509629provisional |
| 4 attempts | total | $0.17810231partly provisional |
All the GPU work
about $0.18
Across four billed GPU attempts, partly provisional.
- The video is an edited progression, not a recording of the generation.
- The refined model was finished by hand in local Blender. That part was not GPU compute.
- It is not a speed or savings claim. Your job gets its own estimate.
04 / For developers
The same flow.
Estimate, then run.
Estimating never spends. Running needs your approval and prepaid credit.
STEP A · ESTIMATE FIRST · FREE
Price it. Book nothing.
No spending limit needed. Nothing is booked and nothing is charged.
estimateExecution({ spec })
→ likely cost, its range, runtime limit,
and what is still missingSTEP B · RUN AFTER APPROVAL
Your limit, then the GPU.
You choose budgetUsd. Your agent submits only after you approve.
estimateExecution({ spec + budgetUsd }) → specDigest
submitJob({ spec, specDigest }) → executionId
getStatus({ executionId }) → stage, nextAction
getArtifact({ executionId }) → meshes + sha256Another demo
A 3D render, three frames.
A second, smaller proof: a Blender Cycles benchmark.
Blender benchmark · 2026-10-07
The command center: 3 of 540 frames.
An agent asked for timeline frames 1, 270 and 540 of an 18-second shot. They came back as 1920×1080 PNGs, together with the editable .blend. A 3-frame benchmark. The full 540-frame film has not been rendered.
Ask your agent for an estimate.
Connect it to Symbioza and ask what your 3D job would cost. An estimate is free and books no compute.