GPT-6 Astra controls robots surprisingly well, but is still too slow to do without specialised models

After briefly topping the RoboDojo benchmark, OpenAI’s general-purpose model has improved the success rate of a block-placement task from 5 per cent to 95 per cent. Kai Williams plays down the significance of this.

In his letter Understanding Robots, Kai Williams looked at OpenAI's statement when GPT-6 Astra was released on September 3: the model would be good at controlling computers, but also robots. The results are striking. Astra briefly took first place in the RoboDojo benchmark, ahead of models designed for robotics such as π0.5 and MolmoAct2. On a block placement task, the success rate jumped by 5% with Fable 5 to 95% with Astra.

However, the author considers this ranking to be too flattering: real-world testing was interrupted after material damage, the assessment was done in simulation and Astra has since returned to seventh place. The model remains too slow for most uses, reacts poorly to changes in the environment, cannot turn on board a robot and pains on precise or coordinated movements.

Its conclusion: large generalist models will not replace specialized vision-language-action models, but hybrid systems, where Astra plans and a VLA runs, should surpass each of the two alone.