Benchmark profile

VoxelBench Image-Prompt Leaderboard (VoxelBench Image)

A live human-preference benchmark where multimodal models build voxel structures from image references and voters compare anonymous results produced from the same prompt.

Data verified July 29, 2026

How to read this leaderboard

The board measures which rendered voxel builds voters prefer when models receive an image reference. Higher ratings indicate stronger pairwise results against the tested field, not absolute reconstruction accuracy.

Operator receipt: 14 sourced rows are currently displayable on this page; the leading published row is Gemini 3.1 Pro Preview at 1701.

Honest limit: The source does not publish a fixed, versioned task set or a controlled reconstruction metric. Ratings can move as prompts, models, and voters change, and the complete generation pipeline contributes to every result.

VoxelBench text-prompt leaderboard Multimodal and grounded rankings

Official live leaderboard Blind comparison voting

How we show VoxelBench image-prompt ratings

We mirror the image-prompt table from VoxelBench's public leaderboard API. The source page shows models after at least 50 votes and ranks the eligible rows with Glicko-2; the snapshot keeps each rating, deviation, 95% confidence interval, vote count, win rate, and win/loss/tie record.

Voters compare two anonymous voxel structures produced for the same image-reference prompt. The result mixes model behavior with VoxelBench's generation tools, instructions, settings, prompt mix, renderer, and voter pool, so we keep it display only and separate from weighted model rankings.

14 eligible model rows50+ votes per modelImage-prompt buildsBlind pairwise votesGlicko-2Display only

VoxelBench leaderboard Blind comparison voting VoxelBench prompts

Glicko-2 rating on VoxelBench image-prompt leaderboard — July 29, 2026

BenchLM mirrors the published glicko-2 rating view for VoxelBench image-prompt leaderboard. Gemini 3.1 Pro Preview leads the public snapshot at 1701 , followed by Claude Opus 4.6 (Max) (1460) and Gemini 3 Pro Preview (1454). BenchLM does not use these results to rank models overall.

Gemini 3.1 Pro Preview

Google

1,267 votes · 83.4% win · 95% CI 1587–1815

gemini-3.1-pro-preview

1701

Overall —

Claude Opus 4.6 (Max)

Anthropic

2,044 votes · 67.2% win · 95% CI 1382–1538

claude-opus-4-6

1460

Overall —

Gemini 3 Pro Preview

Google

3,302 votes · 68.7% win · 95% CI 1374–1534

gemini-3-pro-preview-vertex

1454

Overall —

14 modelsExternal benchmark mirrorsCurrentDisplay onlyUpdated July 29, 2026

Glicko-2 rating table (14 models)

Score

Gemini 3.1 Pro PreviewGoogle1,267 votes · 83.4% win · 95% CI 1587–1815

1701

Claude Opus 4.6 (Max)Anthropic2,044 votes · 67.2% win · 95% CI 1382–1538

1460

Gemini 3 Pro PreviewGoogle3,302 votes · 68.7% win · 95% CI 1374–1534

1454

GPT-5.4 (xHigh)OpenAI1,419 votes · 66.8% win · 95% CI 1366–1534

1450

GPT-5.2 (xHigh)OpenAI2,931 votes · 67.4% win · 95% CI 1330–1482

1406

Claude Opus 4.7 (Max)Anthropic2,283 votes · 51.8% win · 95% CI 1234–1382

1308

Claude Opus 4.5 (High)Anthropic3,095 votes · 59.0% win · 95% CI 1226–1378

1302

Gemini 2.5 ProGoogle2,283 votes · 48.1% win · 95% CI 1112–1288

1200

GPT-5 (High)OpenAI2,494 votes · 50.6% win · 95% CI 1110–1282

1196

GPT-5 Codex (High)OpenAI2,547 votes · 46.4% win · 95% CI 1081–1253

1167

GPT-5.1 (High)OpenAI3,266 votes · 42.1% win · 95% CI 1042–1194

1118

Claude Opus 4.1 (64K Thinking)Anthropic2,526 votes · 40.4% win · 95% CI 1001–1173

1087

O3OpenAI2,594 votes · 29.9% win · 95% CI 868–1048

958

Claude Haiku 4.5 (64K Thinking)Anthropic3,006 votes · 21.4% win · 95% CI 778–962

870

The published VoxelBench Image snapshot places Gemini 3.1 Pro Preview first at 1701. The third row is 247 score units behind. The broader top-10 range is 534 score units, so the table still separates the published systems.

14 models have been evaluated on VoxelBench Image. The benchmark falls in the External benchmark mirrors category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. VoxelBench Image is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About VoxelBench Image

Year

2025

Tasks

Live image-reference prompts for 3D voxel construction

Format

Glicko-2 rating from blind pairwise votes

Difficulty

Visual grounding, 3D construction, and aesthetic quality

The mirrored image-prompt view keeps only models with at least 50 votes, matching the public page. Each row preserves the Glicko-2 rating, deviation, confidence interval, vote count, win rate, and win/loss/tie record.

VoxelBench leaderboard Public benchmark source

BenchLM freshness & provenance

Version

Live VoxelBench Glicko-2

Refresh cadence

Rolling

Staleness state

Current

Question availability

Prompts browsable; full task set not versioned

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

FAQ

What does VoxelBench Image measure?

A live human-preference benchmark where multimodal models build voxel structures from image references and voters compare anonymous results produced from the same prompt.

Which model leads the published VoxelBench Image snapshot?

Gemini 3.1 Pro Preview currently leads the published VoxelBench Image snapshot with 1701 glicko-2 rating. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on VoxelBench Image?

14 AI models are included in BenchLM's mirrored VoxelBench Image snapshot, based on the public leaderboard captured on July 29, 2026.

Last updated: July 29, 2026 · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 2,000+ readers.

One email each week. Unsubscribe anytime.