Three AI models, 17 furniture photos, one attempt each.

The test

Each model gets the photo and the instruction “model the furniture in this photo”, with the same system prompt and the same tools the in-app assistant uses: read the model, write it, edit it, zoom into part of the photo, and ask a question. It has to deliver valid furniture: connected, supported, and with no boards running into each other.

The photos show real sheet-goods furniture: bookcases, TV stands, display cabinets, benches and cube shelves. This is an internal test: examples from this set informed our assistant work. It is not an independent held-out test.

What we saw

Terra took less time and used fewer tool calls. Compare rows, openings and proportions in the front views: a valid model is not necessarily a faithful copy of the photo.

Sonnet 5.5 averaged 13.6 seconds per photo. Its total cost was about four times Terra’s in this run.

Haiku 5.5 is the cheapest per photo, but it produces about ten times more output tokens than Terra and takes longer. Most of that output is reasoning, not furniture JSON.

A photo may hide the back, depth or a joint. These models do not prove strength or load capacity: check sizes, material, joints and wall fixings before you build.

Limits of this test

These are arithmetic means over 17 photos, including every request and tool call in each session. Input includes cache reads and writes; output includes reasoning. Models could correct their work with tools, but failed sessions were not restarted. Two unusable references were excluded before these runs.

Cost is the API list-price equivalent, not a bill for the subscriptions used. Terra reported no cache writes: including unreported writes could raise its estimate from $0.18 to about $0.21. Time measures the agent loop, including the network, not a person reviewing the plans.

It is a small set without repeated trials: it does not establish which model is best for every kind of furniture. Download the CSV results · Read the full method.