PhysBrain 1.5 Tops Open-Source Models on 28 Public Benchmarks, Nears GPT-6 Astra

Summary

The report says the Chinese company DeepCybo released the physical foundation model PhysBrain 1.5 on September 9, 2026. It scored 72.5 on 28 public benchmarks, ranking first among participating open-source models, with 14 first-place and 10 second-place open-source results. Its score is within 1 point of GPT-6 Astra (73.3) and Gemini 3.6 Flash (73.0), and above Hy-Embodied-VLM-1.0 (66.0) by about 6.5 points and RynnBrain 1.1 (63.1) by 9.4 points. On individual benchmarks, it scored 73.9 on RoboSpatial-Home, 84.0 on Part-Affordance and 89.8 on RoboRefit. The 2B and 8B weights, technical report and evaluation kit are open-sourced; the 2B version scored 66.6 on the same tests and is described under official rules as reference-only. The report also cites a Robocurve experiment in which GPT-6 Astra completed 19 of 20 block-in-bowl tasks, versus 8 for Claude Fable 5.1, but its success rate on a millimeter-level puzzle-insertion task fell from 95% to 10%.