Claude Fable 5.2 Tops Artificial Analysis's Robot Arm Test at 75. It Cracked the Hardest Task by Reading a Tape Measure It Never Picked Up.
Artificial Analysis posted on November 12 that Claude Fable 5.2 scores 75 on AA-Robotics v0.1, its research-preview benchmark that hands a frontier model control of a six-axis Hiwonder arm through its own lab's coding agent and grades eight zero-shot tabletop tasks on end state. GPT-6 Astra is at 69 and GPT-6.1 Sol at 65. Fable 5.2 is the first model to score on the measure-the-box task, which it solved from video by using the tape measure's markings as a scale rather than handling the tape. The October board had Astra at 69.4, Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8, and Fable 5.1 46.3, with models taking about 30 minutes per task against a human's 20 seconds. Fable 5.2's cost and time per task are not yet published. We read AA's article and posts.
By T.W. Ghost
Artificial Analysis posted on November 12, 2026, that Claude Fable 5.2 scores 75 on AA-Robotics v0.1, ahead of GPT-6 Astra at 69 and GPT-6.1 Sol at 65. AA-Robotics is a research preview AA launched on October 14. A frontier model, running inside its own lab's coding agent (Codex, Claude Code, or Antigravity), gets Python and HTTP control of a six-axis Hiwonder arm, camera views, and eight tabletop tasks it has never seen, and is graded on the end state of each task with partial credit. The headline from the post is the task nobody had solved: measure the box. Fable 5.2 is the first model to score on it, and it did so from video, using the tape measure's printed markings as a scale, rather than by picking the tape up.
That is a clever answer to a question about hands. It is also the whole shape of this board: six models with scores, eight tasks, one arm, and a human judge. The October article that launched the benchmark put Astra at 69.4, Sol at 64.6, Gemini 4 Argon at 55.3, Claude Opus 5.5 at 48.8, and Fable 5.1 at 46.3, and reported that the best models take about 30 minutes per task against a human average of 20 seconds, about 100 times slower, with the cheapest model costing about 10 times a person at the US average wage. Fable 5.2's own cost and time per task were not on the article page when we read it today. We read AA's article and its two posts. Here is what held up.
The Board
| Model | AA-Robotics v0.1 | Measure-box (October) | Cost per task (October) |
|---|---|---|---|
| Claude Fable 5.2 | 75 (Nov 12 post) | First to score (Nov 12 post), via video | Not published |
| GPT-6 Astra | 69 (Nov 12); 69.4 (October) | 0 | $16.36 |
| GPT-6.1 Sol | 65 (Nov 12); 64.6 (October) | 17 | $1.93, cheapest |
| Gemini 4 Argon | 55.3 (October) | 0 | |
| Claude Opus 5.5 | 48.8 (October) | 17 | |
| Claude Fable 5.1 | 46.3 (October) | 0 | |
| Gemini 4 Carbon | No score; 7 h 33 min in the lab, broke a gripper, retry granted |
The win is a workaround. Measure-box was built to test handling a tape measure. Fable 5.2 scored by not handling it, reading the markings off the video instead. AA presents that as resourceful, and it is; it is also a reminder that end-state grading pays for any route to the answer.
Slow and expensive, by AA's own October account. About 30 minutes a task against 20 seconds for a person. Sol was the cheapest model at $1.93 a task and Astra cost $16.36, with 70% of spend going to input context. Fable 5.2 has no cost line here yet; on AA's main Intelligence Index it costs twice what Fable 5.1 does per task.
Hands are still the weak point. AA's October grasp-success graphic: Fable 5.1 about 45% over 95 attempts, Opus 5.5 44%, Argon 40%, Sol 28% over 111. Carbon spent 7 hours 33 minutes in the lab, broke a gripper, and was granted a retry; it has no score yet.
Model plus harness. Each model drives the arm through its own lab's coding agent, so the number scores Claude Code as much as Fable, and Codex as much as Astra.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Fable 5.2 75; Astra 69; Sol 65; first measure-box score, via video and tape markings | Verified (AA post on X) |
| October board: 69.4, 64.6, 55.3, 48.8, 46.3; measure-box 0, 17, 17, 0, 0 | Verified (AA article) |
| About 30 min per task vs 20 s human; about 100x slower; cheapest model about 10x a person at US average wage | Verified (AA article) |
| Sol $1.93 per task; Astra $16.36; 70% of spend on input context | Verified (AA article) |
| Grasp success: Fable 5.1 about 45% (n=95), Opus 5.5 44%, Argon 40%, Sol 28% (n=111) | Verified (AA graphic) |
| Carbon 7 h 33 min, broke a gripper, retried; no score | Verified (AA post on X) |
| Fable 5.2 cost and time per task on AA-Robotics | Not published in anything we read |
| 2 mm tolerance on measure-box; "29-point leap" | Press (AlphaSignal); 29 is 75 minus Fable 5.1's 46, not an AA figure |
Who Should Use It
Use the result as a signal that a frontier model can now run a cheap arm through a coding agent well enough to score 75 of 100 across eight tabletop tasks, with Fable 5.2 at the front.
Skip any production reading; a research preview with six scored models, human grading, half-hour tasks, and no cost line for the winner is a demonstration of capability, not a bill of materials.
The frontier is elsewhere: the day before, Runway put Praxis-1, a purpose-built robot model, out under Apache 2.0, where first place on NVIDIA's board is 46.2% and every number is Runway's. Our model comparison has the current standings side by side.
Sources
- Artificial Analysis: AA-Robotics
- Artificial Analysis on X: Claude Fable 5.2 on AA-Robotics
- Artificial Analysis on X: Gemini 4 Carbon in the robotics lab
- AlphaSignal: Anthropic's Claude Fable 5.2 tops robot arm benchmark (secondary)
*Not sure whether the top of a six-model board tells you anything about your workload? Take the free 2-minute quiz and get matched. Then read what the same model costs on the main Index: Claude Fable 5.2.*