Hard Vision, Easy Vision: What GPT-6 Astra Reveals Across Computer Vision

Hard Vision, Easy Vision. What GPT-6 Astra Reveals Across Computer Vision. By Hanoona Rasheed and five others, from the Mohamed bin Zayed University of Artificial Intelligence and Apertix. Published on the twenty-eighth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Frontier general-purpose systems are moving past visual understanding into capabilities that dedicated computer vision models have traditionally handled. The question this raises for the field is how far that reach extends, and what remains hard. The answer offered is a pattern. Tasks built on semantic interpretation, reasoning and object-centric prediction are increasingly approaching or reaching available reference levels. Tasks that demand metric geometric accuracy, faithful reconstruction, temporally consistent dense prediction, or specialized fine-grained visual knowledge still show larger gaps.

The evidence is a broad evaluation. GPT-6 Astra is tested alongside five other frontier general-purpose systems, across thirty-four capabilities and fifty-five benchmarks, spanning nine areas of computer vision. Their performance is compared with dedicated models, and with humans where suitable references exist. That breadth is the point of the design. A claim about what general systems can and cannot do in vision needs to be tested across the field, not on a handful of favourite tasks, and the reference points give each score something to be measured against beyond the other general systems.

Astra itself shows broad visual capability, with substantial gains over the other frontier systems in visual and spatial reasoning and in several forms of structured prediction. The split between easy and hard, though, is described as consistent across the state-of-the-art systems, not particular to one of them. Where a task asks what something is, or what follows from it, the systems are closing in. Where it asks exactly where something is, or for a faithful, dense and temporally stable reconstruction, they are not yet there.

Some of the gaps can be narrowed. Additional reasoning and specialist tools close selected ones, but their benefit varies from capability to capability, so neither is a general remedy. The comparisons are also bounded by what references are available: dedicated models and humans are used where suitable ones exist, which means not every capability is measured against both. And the picture is a snapshot of systems that the authors describe as rapidly changing.

The result is a map rather than a scoreboard. Increasingly sophisticated visual tasks are reachable through a general-purpose interface. Precise, fidelity-sensitive perception, the kind that has to get the geometry and the detail exactly right, remains the frontier.