StealthMark 1.8.8

Original tests no model could study for.

Physical and spatial reasoning, planning across many steps, and turning ideas into precise visuals. Every score on this page comes with the result behind it.

Results as of 28 October 2026

See the scores See the evidence

Loading map

Compare any two models

Every benchmark side by side, with the results behind each score.

First model

GPT-6 Astra

Second model

DeepSeek-V4.1-Flash-1017

Weights · DeepSeek-V4.1-Flash-1017, Official weights on Hugging Face

First model

Second model

74.7

Benchmark score

59.2

73.4

General Intelligence

56.8

77.0

Visual Intelligence

63.5

Compare on stealthmodels.com

Scores

What the scores mean

Benchmark score: How well a model's abilities carry over to original, demanding tasks in reasoning, planning and visual work, as one score from 0 to 100. A high score means its skills hold up beyond familiar benchmark questions.
General Intelligence: The ability to reason, plan, solve problems and understand how things relate and behave. A high score means it thinks hard problems through, from careful analysis to plans that span many steps.
Visual Intelligence: The ability to understand visual intent and turn ideas, descriptions and plans into coherent, effective visuals. A high score means design ideas come out the way you meant them.

All models Open weightsThe weights can be downloaded. LocalOpen weights that run on a consumer PC.

Model
GPT-6 Astra Inspect	74.7	73.4	77.0
Claude Opus 5.5 Inspect	72.6	71.3	75.0
Claude Sonnet 5.5 Inspect	70.2	67.0	76.1
GPT-6.1 Sol Inspect	67.8	64.4	74.1
Gemini 4 Argon Inspect	67.4	63.6	74.4
Qwen3.8-Max-0902 Inspect	63.3	62.7	64.5
Claude Fable 5.1 Inspect	63.0	60.6	67.1
Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)Router HF card · Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM), Our findings on Hugging Face Inspect	62.6	58.7	69.6
GPT-6.1 Terra Inspect	60.4	58.0	64.9
DeepSeek-V4.1-Flash-1017 Weights · DeepSeek-V4.1-Flash-1017, Official weights on Hugging Face Inspect	59.2	56.8	63.5
Muse Wonder Inspect	57.0	53.1	64.0
Kimi K3.1 Pro Inspect	56.7	53.9	62.0
GPT-5.6 Sol Inspect	54.1	51.9	58.2
GPT-6 Sol Inspect	51.7	46.1	61.6
Qwen3.8-27B Weights · Qwen3.8-27B, Official weights on Hugging Face Inspect	51.7	46.2	61.4
Kimi K3 Weights · Kimi K3, Official weights on Hugging Face Inspect	50.6	44.4	61.4
DeepSeek-V4.1-Flash Weights · DeepSeek-V4.1-Flash, Official weights on Hugging Face Inspect	50.0	48.5	53.0
Claude Haiku 5.5 Inspect	49.5	47.6	52.8
GLM-5.3-Flash Weights · GLM-5.3-Flash, Official weights on Hugging Face GGUF · GGUF weights on Hugging Face HF card · GLM-5.3-Flash, Our findings on Hugging Face Inspect	48.5	42.3	59.4
Gemini 3.8 Flash Inspect	47.9	43.9	55.2
MiMo-V2.6-Pro Weights · MiMo-V2.6-Pro, Official weights on Hugging Face Inspect	47.9	44.3	54.3
Grok 4.7 Inspect	47.2	48.4	44.8
Qwen3.8-Flash Weights · Qwen3.8-Flash, Official weights on Hugging Face Inspect	46.6	38.5	60.3
MiniMax M3.1 Flash Inspect	45.9	35.8	62.6
GPT-5.6 Terra Inspect	45.8	41.9	52.7
MiniMax M3.1 Flash (early version) HF card · MiniMax M3.1 Flash (early version), Our findings on Hugging Face Inspect	45.5	35.7	64.4
GLM-5.3 Weights · GLM-5.3, Official weights on Hugging Face Inspect	45.1	37.6	57.2
MiMo-V2.6-Flash Weights · MiMo-V2.6-Flash, Official weights on Hugging Face Inspect	44.6	41.0	49.8
Muse Spark Open Inspect	44.1	42.6	46.6
Hy4 preview Weights · Hy4 preview, Official weights on Hugging Face Inspect	43.9	41.3	48.6
Mistral Large 4 Inspect	43.3	34.4	58.0
Ling 3.1 Flash Weights · Ling 3.1 Flash, Official weights on Hugging Face Inspect	42.2	35.5	51.2
DeepSeek-V4-Pro-0813 Weights · DeepSeek-V4-Pro-0813, Official weights on Hugging Face Inspect	41.8	41.6	42.3
GPT-6.1 Terra Low Inspect	41.3	34.3	53.1
Muse Spark 1.3 Inspect	39.8	38.3	42.3
GPT-6 Luna Inspect	39.1	31.3	52.0
Step 5 Preview Weights · Step 5 Preview, Official weights on Hugging Face Inspect	39.0	37.1	41.9
GPT-5.6 Luna Inspect	37.8	30.4	50.2
GPT-5.5 Inspect	36.5	36.6	36.3
MiniMax M3 Weights · MiniMax M3, Official weights on Hugging Face Inspect	34.6	28.0	45.3
Qwen3.6-27B Weights · Qwen3.6-27B, Official weights on Hugging Face Inspect	32.4	25.6	43.9
Beam 1.0 501B A23B Weights · Beam 1.0 501B A23B, Official weights on Hugging Face Inspect	31.9	30.9	33.8
Inkling-Small Weights · Inkling-Small, Official weights on Hugging Face Inspect	27.6	25.5	31.1
Gemma 4 31B Weights · Gemma 4 31B, Official weights on Hugging Face Inspect	25.8	19.1	36.7
Qwen3.6-35B-A3B Weights · Qwen3.6-35B-A3B, Official weights on Hugging Face Inspect	20.8	12.7	34.0
Iceman Contributor HF card · Iceman Contributor, Our findings on Hugging Face Inspect	19.2	11.7	31.0
Glyph Cluster HF card · Glyph Cluster, Our findings on Hugging Face Inspect	18.9	12.0	29.8
Nemotron 3 Ultra Weights · Nemotron 3 Ultra, Official weights on Hugging Face Inspect	16.7	12.2	26.5
Gemma 4 26B A4B Weights · Gemma 4 26B A4B, Official weights on Hugging Face Inspect	16.5	9.4	28.9
MAI-Thinking-1 Inspect	14.1	7.4	24.6
Qwen3.5-9B Weights · Qwen3.5-9B, Official weights on Hugging Face Inspect	10.9	7.9	15.6
Partial results
Callstack ApexPartial coverage HF card · Callstack Apex, Our findings on Hugging Face Inspect Partial coverage Not scored in Foresight	49.5	40.4	59.6
Fledge AlphaSuspected routerPartial coverage HF card · Fledge Alpha, Our findings on Hugging Face Inspect Partial coverage Not scored in Foresight, Latch	35.8	27.1	45.2
LongCat-2.0Partial coverage Weights · LongCat-2.0, Official weights on Hugging Face HF card · LongCat-2.0, Our findings on Hugging Face Inspect Partial coverage Not scored in Foresight, Latch	25.5	17.2	34.4
Kolibri-1Partial coverage Weights · Kolibri-1, Official weights on Hugging Face Inspect Partial coverage Not scored in Foresight, Latch	13.6	12.3	15.0

View the full results

See the evidence

Each benchmark below shows its highest and lowest scoring result, so you can judge the gap yourself.

Newton's Mirror: Can a model trace light through a hall of mirrors in its head, with no code to do the maths?

A high score means the physical state, geometry and repeated reflections remain consistent.

Reference

GPT-6 Astra: Newton's Mirror benchmark result — Benchmark score: 70.6 out of 100
Recorded 30 September 2026View result · GPT-6 Astra, Newton's Mirror

MAI-Thinking-1: Newton's Mirror benchmark result — Benchmark score: 8.0 out of 100
Recorded 9 October 2026View result · MAI-Thinking-1, Newton's Mirror

All Newton's Mirror results

Ground Truth: Can a model reason through a scene where every detail depends on another?

A high score means the model follows how placement, materials and light affect one another.

Reference

Claude Opus 5.5: Ground Truth benchmark result — Benchmark score: 80.6 out of 100
Recorded 6 October 2026View result · Claude Opus 5.5, Ground Truth

MAI-Thinking-1: Ground Truth benchmark result — Benchmark score: 1.9 out of 100
Recorded 9 October 2026View result · MAI-Thinking-1, Ground Truth

All Ground Truth results

Latch: Can a model find the best plan, not just a plan, and get better at it level by level?

A high score means the robots coordinate efficiently, avoid irreversible mistakes and improve across levels.

Claude Opus 5.5: Latch, level 5 — Benchmark score: 87.6 out of 100
5 of 5 levels completed
Level 1 · Cleared · 5.2 points
Level 2 · Cleared · 10.2 points
Level 3 · Cleared · 19.2 points
Level 4 · Cleared · 18.2 points
Level 5 · Cleared · 20.6 points
Watch the replay · Claude Opus 5.5, Latch

Glyph Cluster: Latch, level 5 — Benchmark score: 3.6 out of 100
5 of 5 levels completed
Level 1 · Cleared · 5.2 points
Level 2 · Out of turns · 2 points
Level 3 · Finished early · 4.2 points
Level 4 · Out of turns · 1 points
Level 5 · Finished early · 9 points
Watch the replay · Glyph Cluster, Latch

All Latch results

Foresight: Can a model stay one step ahead of the ghosts?

A high score means the model collects points while anticipating where the ghosts will move.

GPT-6 Astra: Foresight, level 3 — Benchmark score: 76.1 out of 100
3 of 3 levels completed
Level 1 · Survived · 27 points
Level 2 · Survived · 11 points
Level 3 · Survived · 18 points
Watch the replay · GPT-6 Astra, Foresight

Iceman Contributor: Foresight, level 3 — Benchmark score: 3.2 out of 100
3 of 3 levels completed
Level 1 · Out of lives · 0 points
Level 2 · Out of lives · 3 points
Level 3 · Out of lives · 1 points
Watch the replay · Iceman Contributor, Foresight

Planning ahead: the two traps in Foresight

Watch on YouTube · Planning ahead: the two traps in Foresight

All Foresight results

Mona Lisa: Can a model rebuild what it can describe perfectly?

A high score means the result preserves familiar proportions, expression, composition and light.

Reference

GPT-6 Astra: Mona Lisa benchmark result — Benchmark score: 66.9 out of 100
Recorded 30 September 2026View result · GPT-6 Astra, Mona Lisa

Nemotron 3 Ultra: Mona Lisa benchmark result — Benchmark score: 0.5 out of 100
Recorded 4 October 2026View result · Nemotron 3 Ultra, Mona Lisa

Visual recall: the Mona Lisa from memory

Watch on YouTube · Visual recall: the Mona Lisa from memory

All Mona Lisa results

Selfie: Can a model turn a character brief into an original, expressive design?

A high score means the character follows its brief with clear anatomy, expression and materials.

Different character briefs

GPT-6 Astra: Selfie benchmark result — Benchmark score: 80.9 out of 100
Recorded 30 September 2026View result · GPT-6 Astra, Selfie

Kolibri-1: Selfie benchmark result — Benchmark score: 22.9 out of 100
Recorded 5 October 2026View result · Kolibri-1, Selfie

All Selfie results

Beach: Can a model compose a whole scene that is both correct and beautiful?

A high score means perspective, scale, reflections and composition work together across the scene.

Claude Sonnet 5.5: Beach benchmark result — Benchmark score: 87.0 out of 100
Recorded 30 September 2026View result · Claude Sonnet 5.5, Beach

Qwen3.5-9B: Beach benchmark result — Benchmark score: 12.8 out of 100
Recorded 5 October 2026View result · Qwen3.5-9B, Beach

All Beach results

Raccoon: Can a model show characters actually doing something together?

A high score means the dog, raccoon and jet ski connect through believable contact, perspective and motion.

Gemini 4 Argon: Raccoon benchmark result — Benchmark score: 81.3 out of 100
Recorded 13 October 2026View result · Gemini 4 Argon, Raccoon

Kolibri-1: Raccoon benchmark result — Benchmark score: 10.0 out of 100
Recorded 5 October 2026View result · Kolibri-1, Raccoon

All Raccoon results

Play it yourself

The same levels and rules the models get. See where your score would place.

Foresight

Can you stay one step ahead of the ghosts?

Play Foresight

Latch

Can you steer two robots without one careless push?

Play Latch

Still in stealth

No company has claimed these models yet. Follow us to see who made them.

Iceman Contributor

Our findings on Hugging Face · Iceman Contributor

Glyph Cluster

Our findings on Hugging Face · Glyph Cluster

Follow Stealth Models on Hugging Face

Iceman Contributor: its results and our clues so far

Watch on YouTube · Iceman Contributor: its results and our clues so far