Test‑time scaling of Opus 5 on robot arms

August 19, 2026

We gave Claude Opus 5 control of a pair of YAM arms and evaluated it on an Inspect Robots task:

“Place the left block on top of the right block”

Opus 5 climbed from 46% (low) to 76% (high). GPT‑5.6 Sol (high) averaged 60%.

Opus 5 controls YAM arms with raw joint targets without relying on classical robotics approaches such as IK solvers or motion planners. No VLAs were involved.

Robotics capabilities of Opus 5 scale with effort levels

2026-08-18T17:06:13.654645 image/svg+xml Matplotlib v3.11.1, https://matplotlib.org/ $5 $10 $20 $50 $100 API cost per run (log) 0 20 40 60 80 100 score Opus (low) Opus (medium) Opus (high) Sol (high) Opus (low) Opus (medium) Opus (high) Sol (high)

API cost per run vs weighted score; bold indicates the condition mean. Opus 5 at fast-mode rates ($10/$50 per MTok); Sol at $5/$30.

Milestones

Each run is scored out of 100 across six weighted milestones.

identifies the left block

1. identifies left block 10 pts

grasps the block

2. grasps block 10 pts

lifts it over the other

3. lifts over other block 20 pts

touches the other block's upper side

4. touches its top 20 pts

stacked and still standing

5. blocks stacked 30 pts

6. accurate final report 10 pts

Milestone funnel

2026-08-18T17:06:13.513521 image/svg+xml Matplotlib v3.11.1, https://matplotlib.org/ identifies left grasps lifts over touches top success (stacked) 0% 20% 40% 60% 80% 100% proportion of runs Opus (low) Opus (medium) Opus (high) Sol (high)

Proportion of runs reaching each rung of the ladder, per effort level. Where conditions tie exactly, the shared node and line show every tied color.

Report

PDF preview unavailable in this browser — download the report.

Download the PDF

All runs

ModelEffortScoreTime (min)TranscriptVideoRerun
Opus 5high1009.2viewwatchdownload
Opus 5high7024.0viewwatchdownload
Opus 5high7026.1viewwatchdownload
Opus 5high7026.9viewwatchdownload
Opus 5high7030.7viewwatchdownload
Opus 5medium1003.8viewwatchdownload
Opus 5medium7013.9viewwatchdownload
Opus 5medium7020.8viewwatchdownload
Opus 5medium6014.8viewwatchdownload
Opus 5medium2015.6viewwatchdownload
Opus 5low608.8viewwatchdownload
Opus 5low609.8viewwatchdownload
Opus 5low5017.0viewwatchdownload
Opus 5low4012.7viewwatchdownload
Opus 5low2030.7viewwatchdownload
GPT-5.6 Solhigh7014.5viewwatchdownload
GPT-5.6 Solhigh7018.9viewwatchdownload
GPT-5.6 Solhigh7022.9viewwatchdownload
GPT-5.6 Solhigh6012.3viewwatchdownload
GPT-5.6 Solhigh3021.7viewwatchdownload

Technical Specifications

ArmsI2RT YAM bimanual arms (two 6‑DoF arms with parallel grippers)
Camerasfixed overhead Intel RealSense D435 + one Intel RealSense D405 on each wrist, 224×224 RGB, images attached to every observation (image horizon 2)
Modelsclaude‑opus‑5 (Anthropic Messages wire, fast mode) and gpt‑5.6‑sol (OpenAI Responses wire), 5 runs per condition
EffortOpus swept low / medium / high via output_config.effort, with thinking on its adaptive default (effort steers its depth); Sol at high reasoning effort
Instruction“Place the left block on top of the right block”
Max turns100 LLM calls per run, 16,000 output tokens per call
Actionsone move per LLM call (action horizon 1); arm speed capped at 10% for safety
Harnessinspect‑robots 0.53.0; collision guardrail off, joint‑effort reporting on, rig‑facts document in the system prompt; full wire capture
Scoringmanual milestone judging from the run frames, transcripts, and videos