OpenAI’s GPT-6 Astra demonstrated superior performance controlling robotic arms in two manipulation tasks, completing the block-to-bowl task 19 out of 20 times, compared to 8 completions by Anthropic’s Fable 5.1 and just 1 by Fable 5, according to openai.robocurve.org. Astra also completed this task faster, averaging 2.5 minutes per trial versus 6.8 minutes for Fable 5.1, while costing less per run at $0.94 compared to $2.12.
The evaluation involved two tasks: picking up a red block and placing it inside a bowl, and inserting a round blue puzzle piece into a matching groove. While Astra excelled in the bowl task, it matched Fable 5.1’s performance on the puzzle task with 2 completions out of 20 trials each, stalling at the final insertion step. The cost per run for the puzzle task was $1.36 for Astra and $2.18 for Fable 5.1. The tests used the same YAM robotic arms and the Inspect Robots agent policy for consistency.
These results highlight Astra’s efficiency and reliability in robotic manipulation compared to its closest competitor, Fable 5.1. Astra’s 2.4 times higher completion rate and 2.3 times lower cost per run on the bowl task underscore advances in AI-driven robotics control. Such improvements are critical as robotics increasingly integrate AI for complex physical tasks, with cost and speed being key factors for practical deployment.
The detailed comparison was published on September 4, 2026, by openai.robocurve.org, providing a benchmark for future AI models in robotic manipulation. Astra’s performance sets a new standard for cost-effective and reliable robotic task completion using language model control.