AI 3D Benchmark: AA-3D Leaderboard Scores and Scene Costs

Artificial Analysis launched AA-3D on October 12, 2026, establishing an artificial intelligence (AI) three-dimensional (3D) benchmark that evaluates how effectively automated models build interactive web scenes. The evaluation tracks 21 models working as autonomous agents inside an offline sandbox to create complete, interactive three.js applications from natural-language prompts.

Results from the Artificial Analysis AA-3D article show Claude Sonnet 5.5 leading the overall leaderboard at 1194 Elo, statistically tied with three other Anthropic and OpenAI models. Generation costs vary substantially across the field, spanning from twenty cents per scene on GPT-6 Luna to more than thirty-two dollars on Claude Fable 5.1.

AA-3D Results: Who Ranks First and What a Scene Costs

On Artificial Analysis's AA-3D benchmark of blind human votes on interactive three.js scenes, Claude Sonnet 5.5 ranks first at 1194 Elo, statistically tied with three other Anthropic and OpenAI models. A scene costs from $0.20 for eighth-place GPT-6 Luna up to $32.08 for Claude Fable 5.1. GPT-6 Sol is the cheapest of the top five at $3.46, so cost per scene rather than rank alone decides which model fits a product.

The leaderboard reflects blind pairwise comparisons scored using the Elo rating system (Elo) with 95 percent confidence intervals (CI). While Anthropic's Claude Sonnet 5.5 holds the numerical peak, its confidence interval overlaps directly with three other leading models from Anthropic and OpenAI. The highest-performing open-weights models trail the leaders by more than 200 Elo points.

The launch article text gives Elo for only four of the 21 models; the other ratings are on the live AA-3D leaderboard. Claude Sonnet 5.5 costs 4.4 times as much as GPT-6 Sol per scene, while Claude Fable 5.1 costs 9.3 times as much. See Claude Sonnet 5.5 benchmarks for wider test evaluations.

AI 3D Benchmark Architecture and Sandbox Evaluation Rules

Artificial Analysis evaluates candidate models by deploying each model as an autonomous programming agent inside an isolated offline sandbox environment. The sandbox equips the model with Blender, the Rapier physics engine, Python, a headless browser, and an asset library licensed under Creative Commons Zero (CC0). Each model receives a written creative brief for an individual scene alongside a structured delivery contract shared identically across all benchmark tasks.

Execution runs operate under strict procedural caps of three hours or 250 iterative turns. During each turn, the agent writes code, renders the scene, inspects screenshots of its work, and revises its implementation. The terminal artifact is a standalone index.html file that executes the interactive 3D scene inside a standard web browser using the three.js JavaScript library.

Artificial Analysis recruits a panel of human judges who perform blind pairwise evaluations of completed scenes generated from identical task prompts. The charts rest on more than 10,000 panel votes. In one section, 1,008 scenes were analysed, excluding 23 that break on the voting page. A public voting demonstration exists on the site, but votes cast on that interactive demo do not alter official leaderboard standings. Artificial Analysis announced AA-3D in two posts on X: first and second.

Token Economics and Scene Generation Expense Discrepancies

Generation expenses per scene diverge dramatically from base model token pricing because autonomous agents resend their full execution history on every revision turn. Across the benchmark test suite, input tokens account for 69 percent of total financial spend, while representing between 97 percent and 99 percent of all processed tokens. Most models read nearly all of that input from the cache.

The financial spread between models spans two orders of magnitude, ranging from twenty cents to over thirty-two dollars per completed scene. Per the AA-3D article, GPT-6 Luna builds all 55 benchmark scenes at $0.20 per scene (see GPT-6 Luna pricing) while finishing eighth. Claude Fable 5.1 costs $32.08 per scene (see Claude Fable 5.1 pricing), making it 9.3 times as expensive as fifth-place GPT-6 Sol. GPT-6 Sol has the lowest generation cost of the top five at $3.46 per scene (see GPT-6 Sol pricing).

Replay runs show cost and token volume diverge. During individual replay runs, GPT-6 Astra assembled its first functional page after six minutes and concluded at 68 minutes, costing $11.62. Claude Sonnet 5.5 produced no visual output until minute 42, continuing its refinement loop until two hours and 58 minutes. Although Claude Sonnet 5.5 generated 6.2 times as many output tokens as Astra, its final run cost of $14.93 was only 1.3 times as high due to lower published token rates.

Statistical Ties and Confidence Intervals on the AI 3D Benchmark

The top four positions on the AA-3D benchmark constitute a statistical tie because their 95 percent confidence intervals overlap with Claude Sonnet 5.5's leading score of 1194 Elo. Artificial Analysis reported that all four top-tier models originate exclusively from Anthropic and OpenAI. Sonnet 5.5 holds the top score, but confidence intervals prevent naming a single winner.

Meta's Muse Spark 1.3 occupies seventh place with a published score of 1035 Elo, positioned directly ahead of eighth-place GPT-6 Luna. The best open-weights models are Kimi K3 (982 Elo) and MiMo-V2.6-Pro (977 Elo). Both trail the leader by over 200 points.

Among open-weights options, Kimi K3 shows high execution spend. Kimi K3 spends 68 times as much per scene as GPT-6 Luna for a similar Elo, per the article. Readers examining frontier model capabilities can see Claude Sonnet 5.5 explained and Claude Fable 5.1 explained.

Execution Behaviors That Separate Stronger Models

Two behaviours track Elo most closely on AA-3D: early visual feedback and the number of interactive controls. The top five models inspect a screenshot after a median of 10 to 13 turns. By turn 20, the top five models view their scene in at least 89 percent of their scenes. In contrast, models ranked between 14th and 20th place delay their initial image review until turns 15 to 40.

Early visual inspection correlates tightly with competitive success, displaying a Spearman rank correlation of 0.82 across the 19 evaluated models possessing multimodal image perception. The top five also begin browser-based testing earlier in the run, reaching a median testing turn of 16 against 25. Autonomous agents allocate 82 percent of their total tool usage time to verifying the web page within a headless browser.

Structural complexity further differentiates top-performing submissions from basic implementations. Top-five scenes display 15 to 20 buttons and controls. The bottom three show only six to eight. This yields a Pearson correlation of 0.90 with Elo rating. Frequent visual checks alone do not ensure high quality: MiniMax-M3 inspected its scene 54 times per run (the highest frequency in the study), yet verified its render by turn 20 in only 7 percent of scenes, finishing in 19th place.

• Early visual feedback: Top-five models view screenshots by turn 20 in at least 89 percent of scenes, showing a 0.82 Spearman correlation with Elo across vision-capable models.
• Interactive controls: Leading models implement a median of 15 to 20 functional buttons, correlating strongly with human preference (Pearson 0.90).

Intelligence Index Correlations and Outliers Across Model Families

Each point on the Artificial Analysis Intelligence Index translates to roughly 18 AA-3D Elo points (Pearson 0.92). This correlation proves substantially stronger than alternative programming benchmarks such as Terminal-Bench 4.0 (0.81) or SciCode (0.76). Several models deviate sharply from this trend.

GPT-6 Luna beats its trend by 141 Elo (see GPT-6 Luna explained), and GPT-6 Sol by 93 (GPT-6 Sol explained), while GPT-6 Astra beats it by 46. GLM-5.3 falls 197 points below trend. It cannot view images and hit the 250-turn limit in 64 percent of scenes. Z AI's vision-equipped GLM-5.3-Flash scored 149 Elo points higher than GLM-5.3 despite possessing a lower Intelligence Index rating.

Domain analysis demonstrates that leading models maintain dominance across varied scene categories, claiming the highest win rates in six of seven subject areas. Across the seven tracked categories (landmarks, machines, natural phenomena, living bodies, games, sports and piloting, space, and stories and fantasy), sample sizes remain modest, ranging from two to 10 scenes per cell with a median of seven. In the living bodies category, sixth-place GPT-6.1 Sol and Claude Sonnet 5.5 were level at the top, less than a point apart, though score spreads under 12 points fall within expected sampling variance.

How Product Builders Should Use AA-3D

Engineering teams building automated three.js scenes must balance scene quality against cost per scene and open-weights needs. Anthropic's Claude Sonnet 5.5 and Claude Fable 5.1 deliver top-tier results, but cost 4.4 times and 9.3 times as much as GPT-6 Sol per scene. High-volume pipelines benefit from GPT-6 Sol. It ranks fifth and costs only $3.46 per scene.

Budget-sensitive applications and background asset generation pipelines can use GPT-6 Luna. At $0.20 per scene, Luna completed all 55 tasks. It used CC0 assets in one-third of runs and placed eighth. Teams requiring open weights must accept a performance penalty exceeding 200 Elo points on Kimi K3 or MiMo-V2.6-Pro, and Kimi K3 spends 68 times as much per scene as GPT-6 Luna for a similar Elo, per the article.

Layer3Labs has not run the AA-3D benchmark or independently executed these three.js scenes, and all metrics reflect published Artificial Analysis results.

• Frontier quality focus: Choose Claude Sonnet 5.5 or Claude Fable 5.1 when interactive scene depth justifies higher per-scene expenditures.
• Cost-effective top tier: Choose GPT-6 Sol at $3.46 per scene as the cheapest of the top five, with Fable 5.1 costing 9.3 times as much.
• High-volume budget pipelines: Choose GPT-6 Luna at $0.20 per scene for building all 55 scenes and top-ten performance at minimal cost.
• Open weights: Choose Kimi K3 or MiMo-V2.6-Pro when data governance requires open weights, accepting an Elo score reduction of over 200 points.