# StealthMark Intelligence Benchmarks (overall score) (StealthMark)

> StealthMark, from Stealth Models, scores a model out of 100 on eight original tasks: four on physical reasoning and multi-step planning, and four on producing a precise picture from a brief. The exact instructions are kept private so models can't be tuned to them.

Canonical page: https://benchlm.ai/benchmarks/stealthmark

- Category: [external](/external)
- Last updated: November 13, 2026

## About StealthMark

- Year: 2026
- Tasks: 8 original tasks: Newton's Mirror, Ground Truth, Latch, Foresight, Mona Lisa, Selfie, Beach and Raccoon
- Format: Private instructions, blind judging, rendered references and playable game levels
- Difficulty: Chained physical reasoning, planning without tools and precise visual construction
- Paper: [StealthMark: The AI Benchmark Explained (Stealth Models)](https://stealthmodels.com/stealthmark)

The General Intelligence half has two physics scenes and two games. In Newton's Mirror the model has to work out the reflections of a spinning sphere in a room of mirrors, and in Ground Truth the light and shadows around a few everyday objects; both are compared with rendered references. Latch and Foresight are games the model plays turn by turn with no code and no tools, across several levels. The Visual Intelligence half is four drawing tasks (Mona Lisa, Selfie, Beach and Raccoon) judged on likeness, anatomy, placement, light and finish. Every result is scored without knowing which model made it, and blank answers, refusals and broken output count as failures. The site publishes each image, the judges' notes and every game replay, along with separate General and Visual scores, which we keep in each row note. A ninth task, Pelican, is shown on the results page but is not part of the score.

StealthMark is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (42 models)

| Rank | Model | Creator | Score | Who ran it |
|------|-------|---------|-------|-----------|
| 1 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 74.6 | Run by StealthMark |
| 2 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 72.6 | Run by StealthMark |
| 3 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Anthropic | 70.2 | Run by StealthMark |
| 4 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 67.8 | Run by StealthMark |
| 5 | [Gemini 4 Argon](/models/gemini-4-argon) | Google | 67.4 | Run by StealthMark |
| 6 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 62.9 | Run by StealthMark |
| 7 | [Qwen3.8 Max](/models/qwen3-8-max) | Alibaba | 62.4 | Run by StealthMark |
| 8 | [GPT-6.1 Terra](/models/gpt-6-1-terra) | OpenAI | 60.4 | Run by StealthMark |
| 9 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | DeepSeek | 59.1 | Run by StealthMark |
| 10 | [Muse Wonder](/models/muse-wonder) | Meta | 57.0 | Run by StealthMark |
| 11 | [Kimi K3.1 Pro](/models/kimi-3-1-pro) | Moonshot AI | 56.7 | Run by StealthMark |
| 12 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 53.7 | Run by StealthMark |
| 13 | [GLM-5.3](/models/glm-5-3) | Z.AI | 52.9 | Run by StealthMark |
| 14 | [GPT-6 Sol](/models/gpt-6-sol) | OpenAI | 52.0 | Run by StealthMark |
| 15 | [Qwen3.8-27B](/models/qwen3-8-27b) | Alibaba | 51.5 | Run by StealthMark |
| 16 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 50.6 | Run by StealthMark |
| 17 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Anthropic | 49.5 | Run by StealthMark |
| 18 | [MiMo-V2.6-Flash](/models/mimo-v2-6-flash) | Xiaomi | 49.3 | Run by StealthMark |
| 19 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google | 47.9 | Run by StealthMark |
| 20 | [MiMo-V2.6-Pro](/models/mimo-v2-6-pro) | Xiaomi | 47.9 | Run by StealthMark |
| 21 | [Grok 4.7](/models/grok-4-7) | xAI | 47.2 | Run by StealthMark |
| 22 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | OpenAI | 45.8 | Run by StealthMark |
| 23 | [MiniMax M3.1 Flash](/models/minimax-m3-1-flash) | MiniMax | 45.4 | Run by StealthMark |
| 24 | [Muse Spark Open](/models/muse-spark-open) | Meta | 44.1 | Run by StealthMark |
| 25 | [Hy4 preview](/models/hy4-preview) | Tencent | 43.9 | Run by StealthMark |
| 26 | [Mistral Large 4](/models/mistral-large-4) | Mistral | 43.3 | Run by StealthMark |
| 27 | [DeepSeek V4 Pro 0813](/models/deepseek-v4-pro-0813) | DeepSeek | 42.8 | Run by StealthMark |
| 28 | [Ling 3.1 Flash](/models/ling-3-1-flash) | InclusionAI | 42.4 | Run by StealthMark |
| 29 | [Muse Spark 1.3](/models/muse-spark-1-3) | Meta | 39.5 | Run by StealthMark |
| 30 | [GPT-6 Luna](/models/gpt-6-luna) | OpenAI | 39.1 | Run by StealthMark |
| 31 | [Step 5 Preview](/models/step-5-preview) | StepFun | 39.0 | Run by StealthMark |
| 32 | [MiniMax M3](/models/minimax-m3) | MiniMax | 38.6 | Run by StealthMark |
| 33 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | OpenAI | 37.8 | Run by StealthMark |
| 34 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 36.5 | Run by StealthMark |
| 35 | [Qwen3.6-27B](/models/qwen3-6-27b) | Alibaba | 32.4 | Run by StealthMark |
| 36 | [Beam 1.0 501B-A23B](/models/reflection-beam) | Reflection AI | 31.9 | Run by StealthMark |
| 37 | [Gemma 4 31B](/models/gemma-4-31b) | Google | 29.5 | Run by StealthMark |
| 38 | [Inkling-Small](/models/inkling-small) | Thinking Machines Lab | 27.5 | Run by StealthMark |
| 39 | [Gemma 4 26B A4B](/models/gemma-4-26b-a4b) | Google | 21.2 | Run by StealthMark |
| 40 | [Qwen3.6-35B-A3B](/models/qwen3-6-35b-a3b) | Alibaba | 20.8 | Run by StealthMark |
| 41 | [Nemotron 3 Ultra](/models/nemotron-3-ultra) | NVIDIA | 16.7 | Run by StealthMark |
| 42 | [MAI-Thinking-1](/models/mai-thinking-1) | Microsoft | 14.1 | Run by StealthMark |

## FAQ

### What does StealthMark measure?

How well a model handles tasks it could not have studied for. Four test physical reasoning and planning, including two games played without code or tools. Four ask the model to draw a scene or character from a brief. The eight results combine into one score out of 100.

### Which model leads StealthMark?

GPT-6 Astra leads StealthMark 1.8.10 with 74.6 out of 100, followed by Claude Opus 5.5 at 72.6 and Claude Sonnet 5.5 at 70.2. The highest-scoring model with downloadable weights is DeepSeek V4.1 Flash 1017 at 59.2, ahead of Kimi K3.1 Pro at 56.7.

### Why are StealthMark's instructions private?

Public benchmark questions end up in training data, and scores can then rise faster than the ability behind them. Stealth Models publishes what each task tests, every image and every game replay, but not the exact prompts. The cost is that nobody else can rerun the benchmark.

## Compare Top Models on StealthMark

- [GPT-6 Astra vs Claude Opus 5.5](/compare/claude-opus-5-5-vs-gpt-6-astra)
- [Claude Opus 5.5 vs Claude Sonnet 5.5](/compare/claude-opus-5-5-vs-claude-sonnet-5-5)
- [Claude Sonnet 5.5 vs GPT-6.1 Sol](/compare/claude-sonnet-5-5-vs-gpt-6-1-sol)
- [GPT-6.1 Sol vs Gemini 4 Argon](/compare/gemini-4-argon-vs-gpt-6-1-sol)