StealthMark Intelligence Benchmarks (overall score) (StealthMark)
We show this table for reference; we do not rank on it.
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.
Benchmark score on StealthMark — November 13, 2026
We compile the StealthMark rows from benchmark-owner or independent runs. GPT-6 Astra leads the table at 74.6, followed by Claude Opus 5.5 (72.6) and Claude Sonnet 5.5 (70.2). We do not use these results to rank models overall.
GPT-6 Astra
OpenAI
Run by StealthMarkClaude Opus 5.5
Anthropic
Run by StealthMarkClaude Sonnet 5.5
Anthropic
Run by StealthMark42 modelsCross-domainCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (42 models)
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAIRun by StealthMark | 74.6 | Run by StealthMark | Not reported | Closed |
| 2 | Claude Opus 5.5AnthropicRun by StealthMark | 72.6 | Run by StealthMark | Not reported | Closed |
| 3 | Claude Sonnet 5.5AnthropicRun by StealthMark | 70.2 | Run by StealthMark | Not reported | Closed |
| 4 | GPT-6.1 SolOpenAIRun by StealthMark | 67.8 | Run by StealthMark | Not reported | Closed |
| 5 | Gemini 4 ArgonGoogleRun by StealthMark | 67.4 | Run by StealthMark | Not reported | Closed |
| 6 | Claude Fable 5.1AnthropicRun by StealthMark | 62.9 | Run by StealthMark | Not reported | Closed |
| 7 | Qwen3.8 MaxAlibabaRun by StealthMark | 62.4 | Run by StealthMark | Not reported | Open |
| 8 | GPT-6.1 TerraOpenAIRun by StealthMark | 60.4 | Run by StealthMark | Not reported | Closed |
| 9 | DeepSeek V4.1 Flash 1017DeepSeekRun by StealthMark | 59.1 | Run by StealthMark | Not reported | Open |
| 10 | Muse WonderMetaRun by StealthMark | 57.0 | Run by StealthMark | Not reported | Closed |
| 11 | Kimi K3.1 ProMoonshot AIRun by StealthMark | 56.7 | Run by StealthMark | Not reported | Open |
| 12 | GPT-5.6 SolOpenAIRun by StealthMark | 53.7 | Run by StealthMark | Not reported | Closed |
| 13 | GLM-5.3Z.AIRun by StealthMark | 52.9 | Run by StealthMark | Not reported | Open |
| 14 | GPT-6 SolOpenAIRun by StealthMark | 52.0 | Run by StealthMark | Not reported | Closed |
| 15 | Qwen3.8-27BAlibabaRun by StealthMark | 51.5 | Run by StealthMark | Not reported | Open |
| 16 | Kimi K3Moonshot AIRun by StealthMark | 50.6 | Run by StealthMark | Not reported | Open |
| 17 | Claude Haiku 5.5AnthropicRun by StealthMark | 49.5 | Run by StealthMark | Not reported | Closed |
| 18 | MiMo-V2.6-FlashXiaomiRun by StealthMark | 49.3 | Run by StealthMark | Not reported | Open |
| 19 | Gemini 3.8 FlashGoogleRun by StealthMark | 47.9 | Run by StealthMark | Not reported | Closed |
| 20 | MiMo-V2.6-ProXiaomiRun by StealthMark | 47.9 | Run by StealthMark | Not reported | Open |
| 21 | Grok 4.7xAIRun by StealthMark | 47.2 | Run by StealthMark | Not reported | Closed |
| 22 | GPT-5.6 TerraOpenAIRun by StealthMark | 45.8 | Run by StealthMark | Not reported | Closed |
| 23 | MiniMax M3.1 FlashMiniMaxRun by StealthMark | 45.4 | Run by StealthMark | Not reported | Closed |
| 24 | Muse Spark OpenMetaRun by StealthMark | 44.1 | Run by StealthMark | Not reported | Open |
| 25 | Hy4 previewTencentRun by StealthMark | 43.9 | Run by StealthMark | Not reported | Open |
| 26 | Mistral Large 4MistralRun by StealthMark | 43.3 | Run by StealthMark | Not reported | Open |
| 27 | DeepSeek V4 Pro 0813DeepSeekRun by StealthMark | 42.8 | Run by StealthMark | Not reported | Open |
| 28 | Ling 3.1 FlashInclusionAIRun by StealthMark | 42.4 | Run by StealthMark | Not reported | Open |
| 29 | Muse Spark 1.3MetaRun by StealthMark | 39.5 | Run by StealthMark | Not reported | Closed |
| 30 | GPT-6 LunaOpenAIRun by StealthMark | 39.1 | Run by StealthMark | Not reported | Closed |
| 31 | Step 5 PreviewStepFunRun by StealthMark | 39.0 | Run by StealthMark | Not reported | Open |
| 32 | MiniMax M3MiniMaxRun by StealthMark | 38.6 | Run by StealthMark | Not reported | Open |
| 33 | GPT-5.6 LunaOpenAIRun by StealthMark | 37.8 | Run by StealthMark | Not reported | Closed |
| 34 | GPT-5.5OpenAIRun by StealthMark | 36.5 | Run by StealthMark | Not reported | Closed |
| 35 | Qwen3.6-27BAlibabaRun by StealthMark | 32.4 | Run by StealthMark | Not reported | Open |
| 36 | Beam 1.0 501B-A23BReflection AIRun by StealthMark | 31.9 | Run by StealthMark | Not reported | Open |
| 37 | Gemma 4 31BGoogleRun by StealthMark | 29.5 | Run by StealthMark | Not reported | Open |
| 38 | Inkling-SmallThinking Machines LabRun by StealthMark | 27.5 | Run by StealthMark | Not reported | Open |
| 39 | Gemma 4 26B A4BGoogleRun by StealthMark | 21.2 | Run by StealthMark | Not reported | Open |
| 40 | Qwen3.6-35B-A3BAlibabaRun by StealthMark | 20.8 | Run by StealthMark | Not reported | Open |
| 41 | Nemotron 3 UltraNVIDIARun by StealthMark | 16.7 | Run by StealthMark | Not reported | Open |
| 42 | MAI-Thinking-1MicrosoftRun by StealthMark | 14.1 | Run by StealthMark | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
How to read this leaderboard
Higher is better. GPT-6 Astra leads version 1.8.10 at 74.6, ahead of Claude Opus 5.5 at 72.6 and Claude Sonnet 5.5 at 70.2. The two halves can pull apart: Claude Sonnet 5.5's Visual score (76.1) is close to Astra's (76.9) while its General score is six points lower, and MiniMax M3.1 Flash pairs a Visual score of 62.6 with a General score of 35.8. If you care about one half, read that number in the row note.
Operator receipt: 42 sourced rows are currently displayable on this page; the leading published row is GPT-6 Astra at 74.6.
Honest limit: The instructions are private, so nobody outside Stealth Models can rerun the benchmark or check how the eight tasks are weighted into one number. Much of the scoring is human judgment against a rubric. The score is tied to a version and will move as tasks are added (Latch and Foresight joined on October 19, 2026). We leave out models tested under a stealth name, including ones the site has since identified, and variants we don't list. We show it as display-only; it does not enter overall or category rankings.
Among the reported StealthMark rows, GPT-6 Astra is first at 74.6. The third row is 4.4 score units behind. The broader top-10 range is 17.6 score units, so the table still separates the published systems.
42 models have been evaluated on StealthMark. The benchmark falls in the Cross-domain category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. StealthMark is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
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
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.
Freshness and provenance
Version
StealthMark 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
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.
Questions
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
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.