Artificial Analysis: Claude Fable 5.2 Takes the Lead on Robot-Arm Benchmark AA-Robotics v0.1 With 75, Ahead of GPT-6 Astra's 69, and Is the First Model to Score on the Measure-the-Box Task

Background and Industry Pain Point

Large models can write code, but there has been little comparable public measurement of whether they can understand the physical world, plan actions and actually carry them out. Artificial Analysis launched AA-Robotics v0.1 as a research preview on October 14, scoring frontier models on the same robot arm and the same tasks. At launch GPT-6 Astra led with 69.4 and Claude models placed only fourth and fifth. Claude Fable 5.2, which Anthropic released on November 9, has now taken the top spot.

Architecture and Mechanism

According to AA's methodology, each model runs zero-shot inside a coding-agent harness (Claude in Claude Code, GPT models in Codex), controls a Hiwonder NexArm 6-axis arm through a simple HTTP API, watches the scene through four cameras and writes its own code to plan movements. The eight tasks include pressing a button, placing a toy, pouring a drink, building a tower and measuring a box, and a human judge scores them with partial credit. AA says that on the measure-the-box task Fable 5.2 used video alone: it left the tape measure where it was set up and used its markings as a scale to work out the box's size.

Benchmarks and Measured Results

All figures are from AA. Fable 5.2 scores 75, GPT-6 Astra 69 and GPT-6.1 Sol 65. On the launch leaderboard Gemini 4 Argon scored 55.3, Claude Opus 5.5 48.8 and the previous Claude Fable 5.1 46.3, so Fable 5.2 is almost 29 points ahead of its predecessor. At launch no model scored on the measure-the-box task. AA has not yet published cost per task, time per task or grasp success for this run, and at the time of writing AA's article page still showed the old leaderboard.

Developer Guide

If you are prototyping robotics or embodied agents, this result suggests the general coding agent plus cameras plus a simple control API approach is still improving quickly, and that models will use reference objects in the scene on their own. Keep in mind that at launch models averaged about 30 minutes per task versus about 20 seconds for a human, so this is far from real-time control. Watch the task videos in the run explorer on AA's AA-Robotics article, and wait for AA to publish Fable 5.2 cost data before judging value (Fable 5.2 API list price is $8 per million input tokens and $40 per million output tokens).