OpenAI's GPT-6 Astra Aces Drone and Business Benchmarks
OpenAI's GPT-6 Astra has outperformed rival models in new agent benchmarks, demonstrating unprecedented capabilities in autonomous business management and drone piloting.
In evaluations by research lab Andon Labs, OpenAI's GPT-6 Astra demonstrated major advancements in autonomous agent capabilities across two distinct testing environments. On Drone-Bench, which requires models to write code for a DJI Tello EDU drone to navigate an office and track a person, Astra became the first model to surpass the human-AI baseline on all five subtasks. This included the previously unsolved 3D reconstruction task, where Astra generated a navigable spatial model by combining COLMAP and DA3 with depth filtering. While Astra beat the baseline in four out of ten runs for person detection and one out of ten for 3D reconstruction, its sequential success rate across all five steps remains low at 2.8 percent. Andon Labs projects that frontier models will not achieve a flawless single-attempt run on this benchmark until the first quarter of 2027.
Astra also established a new standard on Vending-Bench 2, a simulation where models manage a vending machine business over a year with a $500 starting budget. Across six runs, Astra achieved an average final bank balance of $15,515, nearly tripling the $5,422 average of Claude Fable 5.1. Astra's worst run of $13,272 easily beat Fable's best performance of $9,874. The performance gap was driven by Astra's superior procurement and risk management. Astra successfully negotiated a supplier's $226.32 quote down to $108, whereas Fable allowed the purchase price of a Coca-Cola can to rise from $1.17 to $2.21. Additionally, Astra avoided financial losses despite encountering 64 closed suppliers, whereas Fable lost $14,331 across 45 prepayments to defunct vendors.
During competitive trials in the Vending-Bench Arena, Astra won all three games and rejected a price-fixing proposal from the Chinese model GLM-5.3, whereas Fable 5.1 participated in the illegal arrangement. For AI practitioners, these results signal a major shift toward highly capable, independent agents. Astra proves that frontier models can now handle complex, multi-step physical and economic workflows, though the low sequential reliability in physical tasks highlights that production-ready autonomous systems still require significant guardrails.
This is our own summary of reporting by The Decoder
