SplatBench
A model rebuilds an object in 3D from eight photos by placing up to 300 Gaussians itself through tool calls, with no optimizer. Half of the 30 tasks start from an empty scene, and half ask it to reassemble a scrambled fit. SplatScore compares renders from eight angles the model never sees, with an empty scene at 0 and gradient descent at 100.
Scores can be negative, because misplaced Gaussians cover background that an empty scene gets right. Each model ran every task twice at medium effort. A single Gaussian in the object's average color scores 4.2, and every model stays below gradient descent with 64 Gaussians, which scores 67.7. Two scanned scenes are for non-commercial research only. There is no human baseline yet.
- Made by
- Perpetual Research
- Released
- October 2026
- Kind
- Spatial reasoning and counting
- Level
- University
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Strongest model tested
- GPT-6 Astra #3 on the capability index
- Last model added
- October 2026
- Scores refreshed
- November 12th, 2026
GPT-6 Astra leads with 49.6
Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 1 GPT-6 Astra medium effort 49.6
- 2 GPT-6.1 Sol medium effort 38.6
- 3 Claude Opus 5.5 medium effort 25.0
- 4 GPT-6 Luna medium effort 8.60
- 5 Claude Sonnet 5.5 medium effort 5.70
- 6 Claude Haiku 5.5 medium effort 3.60
- 7 Grok 4.7 medium effort 2.10
- 8 GPT-5.6 Terra medium effort 1.20
Source: Official SplatBench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 49.6 | medium effort | Sep 2026 | Official leaderboard |
| 2 | GPT-6.1 Sol | 38.6 | medium effort | Sep 2026 | Official leaderboard |
| 3 | Claude Opus 5.5 | 25.0 | medium effort | Sep 2026 | Official leaderboard |
| 4 | GPT-6 Luna | 8.60 | medium effort | Sep 2026 | Official leaderboard |
| 5 | Claude Sonnet 5.5 | 5.70 | medium effort | Sep 2026 | Official leaderboard |
| 6 | Claude Haiku 5.5 | 3.60 | medium effort | Oct 2026 | Official leaderboard |
| 7 | Grok 4.7 | 2.10 | medium effort | Sep 2026 | Official leaderboard |
| 8 | GPT-5.6 Terra | 1.20 | medium effort | Jul 2026 | Official leaderboard |
The best score, 49.6, costs $1.69 a task
Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official SplatBench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-6 Astra | medium effort | 49.6 | $1.69 | ||
| GPT-6.1 Sol | medium effort | 38.6 | $0.26 | ||
| Claude Opus 5.5 | medium effort | 25.0 | $0.71 | ||
| GPT-6 Luna | medium effort | 8.60 | $0.02 | ||
| Claude Sonnet 5.5 | medium effort | 5.70 | $0.04 | ||
| Claude Haiku 5.5 | medium effort | 3.60 | $0.01 | ||
| Grok 4.7 | medium effort | 2.10 | $1.38 | ||
| GPT-5.6 Terra | medium effort | 1.20 | $0.19 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official SplatBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official SplatBench leaderboard
No lab model card reports it yet
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. GPT-5.1 85.4%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7% - MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%
- ChartQA In use Models answer human-written and generated questions about chart images that need visual and arithmetic reasoning. Scored by relaxed accuracy with a 5 percent numeric tolerance.
Command A Vision 90.9%