Artificial Analysis finds Opus 5.5 acted 22 minutes after a robot task passed

• Claude Opus 5.5 reached a passing state after 23 minutes, then kept moving the toy for another 22, according to ArtificialAnlys.
• Only GPT-6.1 Sol and Gemini 4 Argon stopped within a minute of success in the five-model comparison.
• Models decide when to submit for grading in the AA-Robotics experiment shared in the follow-up.

The benchmark’s mistake wall records three occasions when Opus’s extra adjustments pushed the crocodile back over the rectangle’s far edge. Its cost breakdown attributes 70% of model spending to input context.

The pass criterion was simple: the toy crocodile had to sit entirely inside a marked rectangle. Opus 5.5 reached that state at 23:22 and recognized it: “The crocodile looks fully inside now.”

Submission came at roughly 45 minutes, with the toy almost exactly where it had been when it first passed. Nearly half the run became a victory lap.

Artificial Analysis compared the delay between first reaching a passing state and finishing the crocodile attempt:

• GPT-6.1 Sol: within one minute.
• Gemini 4 Argon: within one minute.
• GPT-6 Astra: three minutes.
• Claude Fable 5.1: seven minutes.
• Claude Opus 5.5: 22 minutes.

All five received 100/100 for crocodile placement in the published results. Their total completion times varied substantially: Sol declared completion after five minutes, compared with 45 for Opus.

The AA-Robotics setup places models in isolated containers where they can write and execute code. The experiment supplies:

• Hardware: a Hiwonder NexArm with five joints and a gripper. Models receive the arm’s name without further specifications.
• Control: a simple HTTP API for sending commands.
• Observation: four cameras, three around the table and one on the wrist.
• Submission: instructions for declaring an attempt complete.
• Harnesses: Codex for GPT models, Claude Code for Claude, and Antigravity for Gemini, all at high effort.

Attempts are zero-shot. Once the model submits, the end state is graded against the task’s pass criteria, with 100 points for completion and partial credit for incomplete results.

The reported minutes include reasoning and observation, as well as physical movement. Artificial Analysis says that after the first action, models spend most of their time planning, with relatively short movements between longer planning steps.

Across the eight tabletop tasks, four models averaged roughly 30 minutes per task; Gemini 4 Argon averaged 52 minutes. The human baseline averaged 20 seconds.

GPT-6.1 Sol built a passing tower after about 17 minutes, then continued and ultimately failed the task, according to Artificial Analysis’s run analysis.

Sol received 0/100 for tower building, even while ranking second overall across the eight tasks at 64.6/100.