GPT-6 Astra Outperforms Rivals in Robotic Manipulation Tasks
OpenAI's model shows significant improvements in completion rates and cost efficiency.
1Benchmark Setup
OpenAI's GPT-6 Astra was tested against Claude Fable 5 and 5.1 models using identical YAM robotic arms under the same Inspect Robots agent policy. The evaluation involved two distinct manipulation tasks: placing a red block in a bowl and inserting a puzzle piece into a groove. Each model completed 20 trials per task with human graders scoring performance based on predefined stages of completion.
2Bowl Task Dominance
GPT-6 Astra achieved remarkable success in the block-to-bowl task, completing 19 of 20 trials compared to Fable 5.1's 8 and Fable 5's single success. Astra demonstrated superior efficiency with 2.5 minutes per trial versus 6.8 minutes for Fable 5.1, while operating at nearly half the cost ($0.94 vs $2.12 per run). This represents a 2.4× higher completion rate at 2.3× lower cost.
3Puzzle Task Results
The puzzle insertion task proved more challenging for all models. Both GPT-6 Astra and Fable 5.1 succeeded only twice in 20 attempts, stalling at the final placement step. However, Astra maintained cost advantages at $1.36 per run versus $2.18 for Fable 5.1. Detailed stage analysis shows Astra achieved more consistent intermediate progress despite similar final success rates.
4Performance Metrics
Trials were scored on a five-stage rubric from no approach to successful placement. In the bowl task, Astra reached the final placement stage in 95% of trials versus Fable 5.1's 40% and Fable 5's 5%. For puzzle insertion, all models showed similar stage distributions, with most failures occurring during final positioning despite successful earlier stages.
5Cost Efficiency
GPT-6 Astra demonstrated significant cost advantages across both tasks. In the bowl task, Astra operated at 44% of Fable 5.1's cost per run. Even in the equally successful puzzle task, Astra maintained a 1.6× cost advantage. These economics could make Astra more viable for scaled robotic applications where cost-per-operation is critical.
6Temporal Analysis
Beyond completion rates, GPT-6 Astra showed substantial time savings. The 2.5-minute average trial duration for the bowl task represents a 63% reduction compared to Fable 5.1's 6.8 minutes. Faster cycle times combined with higher success rates suggest Astra could enable more productive robotic workflows in real-world applications.
7Technical Implications
The results suggest GPT-6 Astra has made particular progress in spatial reasoning and object manipulation fundamentals. Its strong performance on the bowl task indicates improved grasp planning and trajectory optimization, while the puzzle task challenges reveal remaining limitations in fine manipulation and force control.
8Industry Impact
These benchmarks demonstrate meaningful progress in AI-powered robotics. GPT-6 Astra's combination of higher success rates, faster operation, and lower costs could accelerate adoption in manufacturing, logistics, and service robotics. The remaining challenges in precision tasks highlight areas for future research and development focus.
Original article: Hacker News — GPT-6 Astra Outperforms Rivals in Robotic Manipulation Tasks