Part 4 of 7|April 2026|5 min

The $0.04 Assembly

EconomicsPipeline

The numbers

Old: one model, sequential. 25 seconds, $0.30, blank screen.

New: five steps, three models. First object in 3 seconds, $0.04 total.


The pipeline

Compose Workflow
reasoningfast/cheapdeterministic
#StepModelLatencyCost
1
Plan
Sonnet
1.5s$0.015
2
Flatten
Deterministic
<0.1s$0
3
Fabricate ×N
Haiku (concurrent:10)
0.5s$0.005
4
Regroup
Deterministic
<0.1s$0
5
Refine
Sonnet
2s$0.010
6
Light
Sonnet
0.8s$0.008
7
Camera
Pure JS
<1ms$0
Total
~6s end-to-end·$0.038

Watch it build

NDJSON Stream

Click "Compose" to start the pipeline

Simulated compose pipeline. In production, each line is a real NDJSON event from Mastra workflows via Sonnet (plan, refine) and Haiku (fabricate, light).

Streaming NDJSON. Real GLB parts. Phase badges showing progress.


The concurrency trick

Sonnet sequential: 12 × 1s = 12s. Haiku parallel: concurrent:10 = 0.5s. Haiku is 20× cheaper and we run 10 in parallel.


The refinement tax

We tried skipping it. 30% of scenes had visible errors. $0.015 is the cheapest QA engineer you'll ever hire.


Cost and latency

Cost & Latency: Before vs After
COST (USD)Old$0.30New$0.038↓ 87% cheaperLATENCY (seconds)Old25sNew4.8sFirst object at 3s↓ 76% faster
Plan
Fabricate
Refine
Light
Old (single call)