GPT-6 Astra's developer launch — and its first robot-arm benchmark

OpenAI put out an introductory video pitching GPT-6 Astra to developers, leaning on the model's jump in 3D-scene generation — single-prompt renderings of gardens, shipyards, cityscapes, even Dyson spheres. Separately, independent researchers at Robocurve ran Astra through the same bimanual robot-arm test they've already used on Anthropic's Fable models: on a simple pick-and-place task Astra succeeded 19 of 20 tries versus Fable 5.1's 8 of 20, at roughly half the cost per run, though both models tied at 2 of 20 on a harder precision-insertion task. Together it's an early signal that Astra's gains reach beyond chat and coding into physical-world manipulation — with the harder task showing real limits still remain.