GPT-6.2 Astra Tops Claude Fable 5.2 in an Interactive 3D Benchmark
GPT-6.2 Astra leads AA-3D. Open its playable demos, understand the score, and use a Three.js brief to test whether the advantage helps your work.

GPT-6.2 Astra placed first on Artificial Analysis's test of interactive browser scenes. On the leaderboard checked November 18, 2026, Astra at maximum reasoning scored 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259. Artificial Analysis's AA-3D leaderboard
The important part is what people judged. They compared interactive scenes produced for the same brief, with the model names hidden. That is a useful positive signal for work where the deliverable needs to invite exploration, communicate a spatial idea and respond to input. It gives frontend teams a specific reason to try Astra on a prototype.
Try one round. Open the ocean-themed pinball demo, launch a ball and change the camera. Then switch to Claude Fable 5.2 on the same task. See which version makes it easier to find the controls and follow the ball. That quick comparison gives the leaderboard a concrete meaning for anyone building a browser prototype.
Developers Codex inspected the published scene pages and their visible controls. The ranking and model runs come from Artificial Analysis. Physics accuracy and performance would need separate tests.
GPT-6.2 Astra vs Claude Fable 5.2: what the scores mean
AA-3D covers 55 briefs across seven domains. Models turn those briefs into browser experiences using three.js. The ranked result comes from human preference comparisons, so 1,318 is an Elo rating, not a percentage of tasks completed correctly.
The leaderboard lists these configurations and uncertainty intervals:
| Configuration | Elo | Reported 95% interval |
|---|---|---|
| GPT-6.2 Astra, Max | 1,318 | 1,286–1,349 |
| Claude Fable 5.2, Max, Default Fallback | 1,271 | 1,246–1,297 |
| GPT-6.2 Sol, Max | 1,259 | 1,231–1,288 |
Astra holds the highest measured score. The intervals overlap; overlapping individual intervals alone do not settle the statistical significance of a difference. A 47-point Elo lead does not mean the experience is 47% better, and it cannot promise a win on your next brief.
Keep Fable's Default Fallback label when citing the comparison. Anthropic's fallback documentation describes retrying a refused request on another model. The label identifies a configuration with that option enabled; it does not tell us which AA-3D scenes used it. We have not established a scene-level fallback count.
Our reading is that this result supports a focused trial for visual work. A useful prototype combines composition, geometry, interface decisions and interaction. A model can produce valid JavaScript while leaving the user with a confusing object floating in an empty canvas. A preference test asks whether the assembled experience works well enough that somebody would choose it over another version.
That matters when the next decision is whether an idea deserves another day of engineering. A navigable concept can expose a bad camera, unclear controls or an unhelpful spatial metaphor before a team commits to the implementation.
Where to try GPT-6.2 Astra's actual 3D demos
The AA-3D Explorer maps individual briefs to each model's generated scene. The two examples below are Astra outputs at maximum reasoning. Open the page, select a task and select GPT-6.2 Astra (max). The direct links below select both; check the displayed task and model if a shared link loses its selection. Start with the visible controls, then try the interactions that would matter for a similar prototype.
For the playable-pinball brief, Astra's output is High Tide—Undertow Arcade, an ocean-themed pinball table. The page includes a scoreboard, flippers, hold-and-release launch instructions and a camera shortcut.
That makes it a useful review exercise. Try a weak launch and a strong launch. Check whether changing the camera leaves the input understandable. Resize the page, then see whether the playfield and controls still make sense together. A good-looking table earns attention quickly; predictable controls determine whether the attention lasts.
For the Eiffel Tower brief, Astra produced Tour Eiffel—Structure Atlas. Its page contains floor selections, a night setting, and options to inspect the iron lattice or isolate a leg. Treat the tower dimensions and explanations in the model's output as unverified.
Here the interesting design choice is the relationship between the model and its interface. The controls give a visitor reasons to change the view. When reviewing an educational prototype, ask whether each control reveals something useful and whether the scene keeps its sense of scale as the camera moves. For a client demonstration, those decisions can matter as much as additional polygon detail.
Compare the same task with Fable or Sol using the model selector. Comparing one model's pinball game with another model's landmark is a comparison of briefs as much as models. Keep notes on observable differences: visibility of controls, camera recovery, missing requirements and broken interactions. The presence of a control alone does not establish that it works reliably.
For examples outside the benchmark, OpenAI's official Three.js games showcase includes Escapement and Paris Atlas. Treat those as demonstrations to explore, with different selection conditions from AA-3D.
Is this one-shot 3D generation, a model file, or browser code?
Artificial Analysis's methodology runs models in its Stirrup agent harness, offline with common tools and limits. Each task gets one attempt, up to 250 turns or three hours, to deliver an HTML page and assets. Coding, browser inspection, Blender, Rapier physics and CC0 assets are available.
A single submitted artifact can contain many rounds of work. Calling it a first-response creation would hide the opportunity to inspect and revise. The measured capability includes decisions about where to spend that effort.
Scenes must pass automated running checks and review to enter voting; failures are excluded. Panelists view each eligible scene for at least 30 seconds, and public votes do not affect the ranking. The Elo score therefore describes preference among admitted scenes, rather than the share of all briefs that produced working results. Source: AA-3D methodology.
These controls surround a preference judgment. A preferred anatomy visualization could still contain an incorrect label. An appealing vehicle simulation could still use simplified physics. Define correctness for the product you intend to ship.
The output format also matters when searching for an “AI 3D model generator.” This benchmark concerns browser experiences built with code. It does not establish the quality of a clean exported mesh, animation rig, printable object or CAD assembly. If your pipeline needs glTF export, named parts, editable geometry or a particular coordinate convention, put that requirement into your brief and inspect the exported artifact.
A convincing browser view may explain an idea well while still needing substantial work before becoming an asset in your existing renderer or design tool. Make that handoff requirement explicit before choosing a model.
The useful advantage is a reviewable visual idea
Consider a developer asked to explain how a machine comes apart. A written description can list components, and a static image can show their arrangement. An interactive prototype lets a reviewer discover whether the proposed sequence is understandable: select a part, move it, restore it, then inspect the connection it covered.
That is our practical interpretation of Astra's result. The opportunity is to reach a reviewable visual idea sooner. The next step is to ask a subject specialist whether it explains the right thing and an engineer whether its code can support the next feature.
Artificial Analysis's launch-day assessment also reports a flat Coding Agent Index score compared with GPT-6 Astra at maximum effort. Use the 3D result to plan a visual prototype trial. Use your own builds and regression tests to decide whether to give the model existing production code.
Our analysis of early Astra coding reviews examines the other side of that decision: testers' reports of premature stops, goal drift and delegated work being replaced. Those accounts concern different workflows from AA-3D. A scene that wins a preference comparison can still require integration work; keep that work and its review cost in your acceptance decision.
Our GPT-6.2 migration and developer test checklist covers repository work, blocker reporting and API pricing. For this result, keep the question specific: can Astra produce a more useful interactive prototype for the same effort your team currently spends?
A Three.js prompt that gives you something useful to compare
Use a bounded task where you can recognize correctness. An exploded product view is a useful candidate because selection, camera movement and restoring state all have observable outcomes. Replace the object below with one you actually understand.
This is an original comparison brief, not a prompt we have tested with Astra:
Build a small Three.js browser prototype explaining a fictional desk fan.
Audience: a customer trying to identify the major parts.
Use procedural geometry for five named parts: base, stand, motor housing,
blade assembly, and front guard. Treat the geometry as illustrative;
do not invent engineering specifications or claim physical accuracy.
Required behavior:
- Start with the whole fan visible and controls outside the canvas.
- Select a part from the scene or a keyboard-operable HTML list.
- Highlight the selected part and show its name in HTML.
- Provide an Explode slider and a Reset button.
- Explode moves parts apart along documented directions; Reset restores
original transforms, camera, selection, and slider state.
- Support orbiting on desktop and touch without hiding the controls.
- Keep the object proportionate when the container resizes.
- Show a useful message if rendering cannot start.
Use the supplied repository and its pinned dependencies. Add no external
assets, analytics, or network services. Keep scene setup, interaction,
and teardown understandable to a maintainer.
Deliver run instructions, changed files, checks actually performed,
and known limitations. Stay within this scope.Give both candidates the same starting repository, assets, permissions and time allowance. Record the exact model, provider, agent version and effort setting for each run; matching effort labels do not establish equal work or spending. Save the first deliverable before sending corrective feedback. Otherwise you risk comparing one model's first attempt with another model's fifth revision.
The constraints are deliberate. Procedural geometry reduces differences caused by asset selection. A fictional object avoids turning the trial into a test of unsourced engineering knowledge. A second selection route lets you inspect whether the UI and 3D scene share consistent state. Reset exposes whether the model designed a coherent interaction or only added effects that look plausible in a recording.
How to review generated Three.js code beyond a screenshot
Write acceptance checks before examining either result. These are proposed checks for the desk-fan brief; we have not run them against the AA-3D examples.
| Check | Evidence to collect |
|---|---|
| Select every part through both input routes | Highlight and HTML label agree; keyboard focus stays visible. |
| Explode fully, select a part, then reset | Original transforms and camera return; selection and slider reset together. |
| Resize from desktop to a 360px-wide container | No stretched object, clipped controls or horizontal page overflow. |
| Orbit and select on a real touch device | Gestures remain usable and controls stay reachable. Record device and browser. |
| Open, close and reopen the scene repeatedly | Check for duplicate handlers, extra animation loops and resources retained across teardown. |
| Force renderer initialization to fail | A readable recovery message replaces a silent blank area. |
Inspect how resizing is implemented. The Three.js responsive-design manual distinguishes the canvas's display size from its drawing buffer, and explains updating camera aspect and the projection matrix. Stretching the canvas with CSS alone can leave a distorted scene. Ask the reviewer to explain how the code coordinates these sizes.
Inspect resource lifetime too. Three.js requires explicit disposal of resources such as geometries, materials and textures when they are no longer needed. A prototype that works on its first mount can still be a poor fit for a route or dialog that users open repeatedly. Review teardown alongside integration, including event listeners and any animation loop the prototype starts.
Choose a performance target on hardware your users actually have before starting the comparison. Record loading time and responsiveness there, with the same viewport and settings. We have no measurements supporting a universal frame-rate promise for Astra-generated scenes; AA-3D's ranking cannot supply one for your device.
What to do before choosing Astra for 3D work
Run the brief on Astra and your current model. Ask a colleague to compare the resulting scenes with the model names hidden. Have a specialist verify any labels, dimensions or causal claims you plan to publish.
Keep mandatory behavior separate from visual preference. A broken reset fails a requirement even if everyone prefers the lighting. Count total time to acceptance: generation, revisions, code review and human repairs. Include rejected attempts in spending comparisons; a token rate alone cannot price the finished prototype.
AA-3D provides a credible reason to put Astra on the shortlist. Choose it if it reaches a usable, understandable experience with less total work on your brief. Keep the artifacts and review notes so the next model has a real baseline to beat.

