Over the past day, X has been debating whether GPT-6 Astra can work directly as a robot policy. Former OpenAI researcher Will Depue (whose bio lists Sora 1 and 2 and o3/4o post-training) called Astra beating specialized robotics models "deeply surprising", then described it as "the final death of specialized ML". Investor Freda Duan posted that "Astra IS robotics' GPT-3 moment", drawing about 84K views. Third-party evaluations point both ways: in a RoboDojo paper, Astra averages 22.48% success and ranks above all 40 public policies, while an EBench report puts it at 46.7%, third of nine models. This is a discussion of opinions, not an official OpenAI conclusion.
Key Takeaways
- ✓Reach: Will Depue's two posts have about 597 and 566 likes and 59K and 37K views; Freda Duan's "10 takes" has about 455 likes, 559 bookmarks and 84K views (as of 19:10 SGT)
- ✓RoboDojo: over 42 tasks and 2,100 trials, Astra averages 22.48% success (Score 28.97), ranking above all 40 public policies; GPT-5.5 gets 0.88% in the same setup (arXiv 2609.24170, as reported by the paper's authors)
- ✓Counterpoint: in the EBench report, Astra with single-shot in-context learning scores 46.7%, behind OpenWAM-alpha at 55.3% and AMapbot 1 at 48.9% (EBench tech report)
- ✓Real arms: Robocurve measured 19 of 20 successes on a block-into-bowl task on YAM arms at about $0.94 per run, but only 2 of 20 on a puzzle insertion task (Robocurve)
- ✓Weak spots: several evaluations find Astra clearly weaker on high-precision control, dynamic control and bimanual coordination
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
For years, the main robotics approach has been to train dedicated vision-language-action (VLA) models: a general model handles high-level planning while a specialized model handles skills like picking and placing. About a month after GPT-6 Astra shipped, several third-party evaluations found that this general model, with no robotics-specific fine-tuning and outputting actions directly, beats specialized models on some benchmarks. That set off a fast-growing debate on X.
What people are arguing
Former OpenAI researcher Will Depue posted on Oct 31 that the result is "deeply surprising" and might mean transfer increases with scale, or that there has been some fundamental algorithmic shift. He later added that Astra may have transferred a lot from computer use, such as rapidly controlling objects from video input. Investor Freda Duan argued that robotics is becoming just another vertical for frontier labs, and that specialized model companies may be pushed down to low-level skills and motor control. These are personal opinions.
Benchmarks and measurements
The numbers come from third parties and use very different protocols. In a RoboDojo paper, Astra averages 22.48% success and ranks above all 40 public policies, while GPT-5.5 gets 0.88%. In the EBench report it scores 46.7% and ranks third. Another paper reports 81.9% on RoboDojo in a setup that allows in-context demonstrations and longer execution. In Robocurve's real-arm tests, Astra succeeded 19 of 20 times on a simple placement task but only 2 of 20 on a fine puzzle insertion. OpenAI has not published official robotics scores for these benchmarks.
Getting started
Teams who want to try it can first use Astra as a high-level planner in simulation on top of their existing low-level control policy, then decide whether to let it output actions directly. Watch precision, latency and cost per run closely, and add failure recovery and safety limits before any real-robot deployment. Each evaluation uses a different protocol and sample size, so avoid comparing their numbers directly.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.