GPT-6 Astra Leads New Intuitive Visual Benchmark at 53.6% — All Frontier Models Fall 40 Points Behind Humans
Elorian and Scale AI published Humanity’s Sixth Sense (HSS) on October 8, a benchmark designed to measure the visual reasoning humans apply automatically: inferring layout, context, spatial relationships, and implicit intent from images and video without the question making those inferences explicit.
The benchmark contains 522 tasks — 288 static images and 234 short video clips totalling 17.6 hours. Tasks span six sub-dimensions: Time (temporal reasoning), Space, Mind (theory of mind applied to scenes), Context, and the image/video modality split.
Results
| Model | Effort | pass@1 |
|---|---|---|
| Human baseline | — | 93.1% |
| GPT-6 Astra | max | 53.6% |
| GPT-6.1 Sol | max | 46.6% |
| Claude Opus 5.5 | xhigh | 44.6% |
| Gemini 3.8 Flash | high | 41.6% |
| Claude Fable 5.1 | xhigh | 40.8% |
| Gemini 3.7 Flash | high | 39.8% |
| Muse Spark 1.3 | max | 37.4% |
| Claude Fable 5 | max | 34.5% |
| Qwen 3.8 Max | max | 33.0% |
GPT-6 Astra leads by 7 points over GPT-6.1 Sol and 9 points over Claude Opus 5.5. Claude Fable 5.1, which tops LiveBench overall, scores below Opus 5.5 here — 40.8% against 44.6% — a reversal of the usual ordering.
The median model in the benchmark scores 30.9%.
The human gap
The 39.5-point gap between GPT-6 Astra and the human baseline (93.1%) is the main finding. The benchmark’s authors argue HSS isolates a form of visual understanding distinct from object recognition or captioning: the reasoning that lets a person look at a half-eaten meal and infer the diner left in a hurry, or at a cluttered desk and understand which object is in active use.
A control run validates the benchmark is actually measuring visual input, not text reasoning. GPT-6 Astra, run again with the image or video stripped out, scores 6.6%. Without the visual signal, the strongest frontier model is barely above random.
Sub-task patterns
Video is harder than images for every model tested. GPT-6 Astra drops from 55.6% on images to 51.1% on video. Claude Opus 5.5 moves from 51.3% on images to 36.5% on video, a 15-point gap.
The Mind sub-dimension — theory of mind applied to visual scenes — is the weakest category for most models. GPT-6.1 Sol scores 36.4% on Mind against 55.2% on Time, the largest within-model gap in the table.
Context
HSS joins ARC-AGI-3 and HLE as benchmarks where the human-to-frontier gap remains wide enough to be informative. It is the first large-scale benchmark specifically targeting implicit visual reasoning rather than explicit visual question answering. Scale AI and Elorian released the full dataset and paper alongside the launch.