Particle.news

OpenAI’s GPT-6.2 Astra Tops 3D and Robotics Benchmarks

Independent tests show Astra leads AA-3D and AA-Robotics but slow runtimes, notable per-task costs, and weaker spatial reasoning undercut claims of immediate production readiness.

Overview

  • OpenAI released GPT-6.2 Astra on Tuesday, November 17, 2026, and the release included a 1.05 million token context window and roughly 20% lower API pricing for developers.
  • Astra scored 1,318 Elo points to lead the AA-3D benchmark for interactive 3D content generation, and it posted the top average of 79 on Artificial Analysis’s AA-Robotics v0.1 test.
  • In the AA-Robotics v0.1 run Astra used the fewest arm commands of any model tested — 668 total and 92 for the box-cleanup task — which suggested improved planning and command efficiency versus peers.
  • Practical gaps remain: models still take about 30 minutes to complete robot tasks versus roughly 20 seconds for a human and running those trials costs about $2 to $16 per task, and Astra underperformed on some spatial-reasoning tests where Google’s Gemini 4 Carbon led the Blueprint-Bench 2 floor-plan metric.
  • Astra is available to paid ChatGPT tiers and API customers, but independent validation, red-teaming and real-world workflow testing are needed to judge reliability, safety and whether automation costs and runtimes make it viable for production use.