AA-3D Benchmark: Best AI Models for Interactive 3D Scenes by Cost and Quality

The cheapest model and the highest score answer different questions

Source: the official launch cost chart and Elo results. Sonnet’s chart mean is $15.184985…; the $14.93 figure elsewhere in the article describes one example run. Consult the live leaderboard for later results.

These are token-cost estimates for the evaluator’s task setup, reasoning configurations and cache treatment. Input/cache and reasoning/output usage contribute to the cost. They are not flat API prices, hosting budgets, Image3D generation costs or quotes for buying a downloadable 3D file. A different brief, cache hit rate or number of revisions can produce a different bill.

The top four models’ 95% intervals overlap Sonnet’s in the launch results. That supports treating several models as plausible quality candidates, rather than using a small Elo gap as a guarantee. Elo is a relative preference scale: dividing it by dollars does not create a meaningful universal “quality per dollar” score.

What the agent actually builds—and how it is judged

The AA-3D methodology specifies 55 written tasks. Each model works in an offline sandbox with the same instructions, tools and limits. Its output is a browser page—index.html plus assets—not a required GLB or STL export.

1. Build and revise. The agent writes scene code, runs browser checks and can inspect screenshots to guide edits. Tools include Three.js, optional Rapier physics and Blender, plus a supplied CC0 asset library. It has up to 250 turns or three hours.
2. Check eligibility. A headless-browser check and review screen scenes before voting. A working page is a prerequisite, not a high-quality score.
3. Compare anonymously. Recruited panelists interact with two scenes for the same brief without seeing model names. Their preference votes determine Elo and its confidence interval; public demo votes do not enter that rating.

Read this as an evaluation of the assembled experience: visual decisions, controls and implementation working together. A compelling scene may combine procedural geometry and supplied assets. Its score does not tell you how faithfully a separate mesh generator reconstructed an input photograph.

Choose against your delivery requirements

Before comparing model names, write a short acceptance checklist for your own scene. For example: the camera must orbit without getting trapped, two controls must visibly change the scene, the reset button must restore the starting state, and the page must remain usable on your target phone. Add accessibility, licensing and offline requirements if they matter to the project.

A limited prototype budget

Start by reviewing the lower-cost candidate’s official examples against that checklist. A rough visual style may be acceptable when the goal is to test whether an interaction is useful. Escalate only when a specific requirement is missed; a more expensive model is not a substitute for a clear brief.

A polished interactive demonstration

Compare the stronger quality candidates on the same task, with the same controls and stopping rule. Look at the resulting page, not just a hero screenshot. A beautiful first view can hide broken navigation, unreadable labels or a control that does nothing. Keep your own result separate from the benchmark’s population average.

Scene code, a video world and a 3D object are different deliveries

AA-3D addresses A. Image3D’s core image-to-3D workflow addresses C; it does not supply the complete agent-built scene or controllable video-world service in A or B. For the broader representation choice, use the existing guide to objects and 3D worlds.

Consider a product exhibit: a coded scene provides the room, camera and hotspots; an object asset supplies the product’s geometry. Combining them is another integration task. Neither a scene’s preference score nor a convincing video proves that the product mesh has correct dimensions, printable walls or a license for your use.

If you need GLB, OBJ or STL, start with the file handoff

Choose an object workflow when the deliverable is a prop, character or product mesh. Inspect its hidden sides before export, then verify the actual file in the destination application.

• GLB: use the image-to-GLB workflow when you want to inspect an object and carry its available materials into a compatible viewer or editor. GLB can be edited after import; check material compatibility.
• OBJ: read the OBJ export guide. Current Image3D Studio export creates geometry without an accompanying MTL or texture package. A textured preview does not mean the OBJ download contains textures.
• STL: follow the image-to-STL workflow, then inspect scale, wall thickness, disconnected parts and supports in a slicer. Export is not proof of a successful physical print.

If this is the task you need, begin in Image3D Studio. Generated-model downloads require the applicable paid access. Purchasing access does not change an existing mesh’s geometry or turn it into a complete interactive world.