Third-party benchmark: OpenAI's GPT-6 Astra outperforms Claude Fable 5.1 at robotic pick-and-place task, ties on harder puzzle-insertion task

An independent evaluator ran OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 through identical robot-arm manipulation tests using the same YAM hardware and Inspect Robots agent framework, 20 trials per task. On a simple block-into-bowl task, Astra completed 19/20 (95%) versus Fable 5.1's 8/20 (40%) and Fable 5's 1/20 (5%), and did so cheaper ($0.94 vs $2.12 per run) and faster (2.5 vs 6.8 minutes). On a harder puzzle-piece insertion task, both Astra and Fable 5.1 tied at 2/20 (10%), failing at the same final step, though Astra was again cheaper.

This gives developers and enterprises evaluating agentic robot control a rare head-to-head with real numbers rather than vendor claims, useful for model selection on simple grasp-and-place tasks where Astra's advantage looks real and substantial. The shared failure on precision insertion is arguably the more important finding: it suggests a genuine capability ceiling in current models for fine manipulation, not just a Claude-specific weakness, which matters for anyone planning near-term deployment of these models in physical automation.

The "2.4x better" framing is true for exactly one narrow task on one robot rig with 20 trials per condition, a small sample that does not establish general manipulation superiority. This is also an incremental follow-up to an earlier Fable-only report, not a fresh independent study design, and lab pick-and-place results on a single arm setup may not generalise to other tasks, hardware, or real-world conditions.

1. 01Whether independent replications on different robot hardware or task sets confirm Astra's bowl-task advantage or narrow it
2. 02Whether OpenAI or Anthropic respond with their own benchmark data or dispute the methodology
3. 03Whether either model shows improvement on the puzzle-insertion task in future point releases, which would indicate the capability ceiling is moving