Atlas · C05 · Perception
Visual understanding
Making sense of images, scenes, documents, charts and spatial layouts.
AI Reliable score
5.5 / 10
Average adult level · peak 8.5 · trend ↑ · grade B
Why this score
Jagged: on expert multimodal exams top models (90.6% MMMU-Pro) exceed the best human experts on the original MMMU (76-89%), but on spatial geometry the best model scored 58.5% vs humans 96%. Reliable 5.5 averages strong document/chart reading with weak spatial reasoning.
On the human scale
2Beginner. Recognises common objects
4Average adult. Average adult: reads scenes, faces, charts, room layouts ← AI today
6Competent pro. Professional: reads technical drawings, spots defects, interprets domain images
8Expert. Expert: radiologist-level reading, complex spatial reasoning
10Best humans. Best human visual experts
Future warning
At its best, AI already reaches 8.5 on this capability. Peak results are never counted in the index; they show where the water may rise next.
Evidence
MMMU-Pro (Vals AI)B
E05 · 2026-09
- Measures
- Expert-level multimodal (image + text) questions across disciplines
- AI result
- Top: Claude Fable 5.1 90.6%. Benchmark saturated and archived.
- Humans
- Original MMMU human experts 76.2%-88.6% (note: baseline is for MMMU, not MMMU-Pro)
- On our scale
- Expert level on academic visual questions.
vals.ai ↗
MathSpatial-Bench (arXiv 2602.11635 v2)B
E06 · 2026-04
- Measures
- Mathematical spatial reasoning from images
- AI result
- Best model GPT-5 58.5%; most models below 60%.
- Humans
- Humans ~96.3%
- On our scale
- Spatial reasoning below average adult: pulls C05 Reliable down.
arxiv.org ↗
A Definition of AGI (Hendrycks et al., CHC-based)B
E07 · 2025-10
- Measures
- 10 CHC cognitive domains, each worth 10% of an 'AGI score'
- AI result
- GPT-5 total 57%: Knowledge 9, Reading-Writing 10, Maths 10, Reasoning 7, Working memory 4, LT memory storage 0, LT memory retrieval 4, Visual 4, Auditory 6, Speed 3 (out of 10 each).
- Humans
- Well-educated adult (definition of 100%)
- On our scale
- Used as a cross-check on the profile shape, not for absolute levels (single model, dated Oct 2025).
arxiv.org ↗
History
First reading: Atlas v0.1, 5 Nov 2026. Every later change will be listed here with its reason.
Disagree? Send a correction
