Research · Hacker News · September 6, 2026

GPT-6 Astra Dominates Claude in Robot Arm Benchmarks at Half the Cost

Robocurve's benchmark shows GPT-6 Astra completing robotic manipulation tasks at 95% success vs Claude Fable 5.1's 40%, at roughly half the cost per run and 2.7x faster.

Based on reporting by Hacker News — analysis by dalili

A new robotics benchmark from Robocurve puts OpenAI's GPT-6 Astra head-to-head with Anthropic's Claude Fable 5 and 5.1 on physical manipulation tasks — and the results show Astra pulling well ahead. Testing on bimanual YAM robotic arms running the Inspect Robots agent policy, Astra completed a block-in-bowl task in 19 of 20 trials (95%), compared to Fable 5.1's 8 of 20 (40%) and Fable 5's 1 of 20 (5%).

The cost and speed differences are equally striking. Astra averaged $0.94 per run at 2.5 minutes, versus Fable 5.1 at $2.12 and 6.8 minutes. That's a 2.3x cost reduction and 2.7x speed improvement on the task where Astra excels. On the harder puzzle-piece insertion task, all three models struggled — Astra matched Fable 5.1 at 2 of 20 completions but at 1.6x lower cost.

The benchmark used 20 trials per model per task, with human grading on a 5-stage rubric from "no purposeful approach" to "placed in final position." Astra's advantage was most visible in consistent high-stage scoring: it reached the final placement stage in 95% of bowl trials, while Fable 5.1 reached it in only 40%.

The results suggest that for embodied AI tasks, the gap between frontier models can be dramatic — and that cost-efficiency matters as much as raw capability when deploying AI-controlled robotics at scale.

Key takeaways

GPT-6 Astra achieved 95% success on block-in-bowl task vs Claude Fable 5.1's 40%
Astra costs $0.94/run at 2.5 min vs Fable 5.1's $2.12 at 6.8 min — 2.3x cheaper, 2.7x faster
On harder puzzle task, all models struggled — Astra matched Fable 5.1 at lower cost
Benchmark used 20 trials per model with human grading on 5-stage rubric

Why it matters

This is one of the most direct head-to-head comparisons of frontier models on embodied AI tasks, and the margin is unusually large. For the growing robotics-as-a-service market, Astra's combination of higher success rates and lower costs could be decisive. The benchmark also highlights that model choice matters enormously for physical AI — a domain where failure means dropped objects and damaged hardware, not just bad text.