AI Robot Arm Benchmark: AA-Robotics Scores and Costs

Artificial Analysis released AA-Robotics v0.1 on October 14, 2026, a research preview that tests frontier language models on a robot arm. The Artificial Analysis article says the benchmark measures how well frontier language models understand, plan and act in the physical world. Artificial Analysis also announced the release on X.

On eight tabletop tasks with a 6-axis Hiwonder NexArm, GPT-6 Astra scored highest at 69.4 out of 100, followed by GPT-6.1 Sol at 64.6. A task costs from $1.93 (Sol) to $16.36 (Astra). Four of the five models average about 30 minutes per task, and Gemini 4 Argon averages 52.

The article says the models are roughly 100 times slower than a human, and the most cost-effective model costs about ten times as much as a human's task cost. It calls them "far too slow and expensive for any kind of production usage."

AA-Robotics Leaderboard: Who Ranks First and What a Task Costs

On the Artificial Analysis AA-Robotics v0.1 benchmark, GPT-6 Astra ranks first with an overall score of 69.4 out of 100, followed by GPT-6.1 Sol at 64.6, Gemini 4 Argon at 55.3, Claude Opus 5.5 at 48.8, and Claude Fable 5.1 at 46.3. A task costs from $1.93 (GPT-6.1 Sol) to $16.36 (GPT-6 Astra) and four of the five models average about 30 minutes against about 20 seconds for a human, so the article calls them too slow and expensive for production (as of October 2026).

GPT-6 Astra costs more than eight times as much per task as GPT-6.1 Sol ($16.36 against $1.93) for a score of 69.4 against 64.6. Token rates are on the GPT-6 Astra pricing page. The source text read does not include task costs for Gemini 4 Argon, Claude Opus 5.5, or Claude Fable 5.1. Check the Artificial Analysis article for updates.

Gemini 4 Argon is the slowest model at 52 minutes per task. The other four average about 30 minutes.

AI Robot Arm Benchmark Methodology and Setup

GPT models ran in Codex, Claude models in Claude Code, and Gemini 4 Argon in Antigravity, all at high effort. The arm is a Hiwonder NexArm with 5 joints and a gripper. The source text read gives no further specifications. Each model runs in an isolated container with brief instructions, a simple HTTP API to reach the arm, and the ability to write and run code. Tasks are zero-shot.

The setup has four cameras: three around the table and one on the wrist.

A run ends when the model judges the task complete or gives up. A human judge scores the final state against the pass criteria out of 100: completing the task earns 100 and partial progress earns partial points.

Eight Tabletop Tasks and Pass Criteria

Six of the eight tasks involve a grasp. The AA-Robotics Score is the mean score across all eight tasks.

All five models scored 100 on pressing the red button and on placing the crocodile toy in the rectangle. Measuring the box to within 2 mm was the hardest task.

No model scored above 17 on the box measurement task. GPT-6.1 Sol and Claude Opus 5.5 each scored 17, while GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 all scored zero. A related guide covers the AA-3D benchmark.

• Press the red button: Pass: the button flashes.
• Place the crocodile toy in the rectangle: Pass: the toy is entirely within the rectangle.
• Pour the drink into the cup: Pass: the drink is in the cup, none spilled, nothing broken.
• Tidy the blocks into the box: Pass: all four blocks are inside the box and the lid is flush.
• Complete the shape matching challenge: Pass: the blocks match the picture.
• Write hello world: Pass: the writing is legible and inside the black box.
• Build the tallest tower: Pass: a freestanding tower at the maximum achievable height.
• Measure the box: Pass: each dimension is within 2 mm.

Physical Reasoning and Grasp Success Rates

Grasp success rate covers the six grasp-involving tasks and orders the models differently from the leaderboard. GPT-6 Astra leads at 45% (n=95), then Claude Fable 5.1 at 44% (n=89), Claude Opus 5.5 at 40% (n=100), Gemini 4 Argon at 36% (n=75) and GPT-6.1 Sol at 28% (n=111).

GPT-6.1 Sol is second overall (64.6) with the lowest grasp success rate, 28%. GPT-6 Astra is first overall with 45%. The source text read gives no cause for the difference.

The article notes that models struggle to judge task completion in physical space and often keep refining an attempt after reaching a passing state. Its Mistake wall section lists misjudged success, poor physical reasoning, and missed grasps or drops as common failures. On the tower task, GPT-6.1 Sol built a passing tower after about 17 minutes, then failed the task.

Operational Speed and Cost Compared to Human Labor

The article says the models are roughly 100 times slower than a human, a gap it calls nearly two orders of magnitude. A human averages about 20 seconds per task. The most cost-effective model costs about ten times as much as a human's task cost.

Gemini 4 Argon spent an average of 21 minutes observing and reasoning before its first action. Artificial Analysis noted that activity durations for Gemini are estimates because the Antigravity harness does not timestamp individual execution steps.

The article prices a human at $37.81 an hour, the average hourly earnings of US private nonfarm employees (Bureau of Labor Statistics, preliminary September 2026), applied to by-hand task time. GPT-6.1 Sol costs $1.93 per task and GPT-6 Astra $16.36. About 70% of spend goes to context (input).

Production Readiness and Benchmark Limitations

Artificial Analysis says the models are "far too slow and expensive for any kind of production usage."

For a buyer, the practical reading is narrow. This is a research preview of eight tabletop tasks on one arm, run by Artificial Analysis, and the article itself says the models are too slow and expensive for production. The scores rank five models against each other on those tasks. They are not a deployment recommendation, and the source read makes none.

AA-Robotics v0.1 is a research preview with eight tabletop tasks on one arm, a Hiwonder NexArm. Only five models are on the leaderboard in the source read. The source read does not say whether more tasks or models will be added, or when.

The article reports no human AA-Robotics Score, only human time and cost per task. Every score, cost and time here is Artificial Analysis's, as of October 2026. For updates, read the Artificial Analysis AA-Robotics article.