Frontier models' robot arm control ability doubles every 54 days

• Frontier multimodal LMs' ability to control robot arms is doubling every 53.9 days (95% CI: 36.1-59.5 days) based on data from July to October 2026.
• Execution speed increased 4.4x, rising from 12.4 subtasks/hour for GPT-5.6 Sol (H) to 54.4 subtasks/hour for GPT-6 Astra (H+UF), while execution cost decreased 6.2x from $2.36/subtask to $0.38/subtask.
• Top models (GPT-6 Astra, GPT-6.1 Sol, Opus 5.5, Gemini 3.8 Flash, Gemini 4 Argon) achieved average task scores of 45-48%, representing consistent partial progress rather than end-to-end success.
• Performance is currently limited by efficiency rather than reasoning capability, with failures mostly due to physical errors like grip failures or distance errors.

Methodology and Models

• Evaluated 9 frontier multimodal LMs across 15 configurations on the I2RT YAM bimanual robot using RoboDojo's 6 tasks.
• Models included OpenAI (GPT-5.6 Sol, GPT-6 Sol, GPT-6 Astra, GPT-6.1 Sol), Anthropic (Opus 5, Opus 5.5), and Google DeepMind (Gemini 3.6 Flash, Gemini 3.8 Flash, Gemini 4 Argon).
• Scoring was based on partial progress (subtasks) determined by GPT-6 Astra (H) analyzing camera video, with 20 trials per model-configuration-task combination.

Performance and Trade-offs

• Top models demonstrated superior efficiency by recognizing and compensating for physical system defects such as arm sag and camera parallax.
• A clear trade-off exists between performance, cost, and time; no single model dominates the Pareto frontier.
• GPT-6 Astra (H+UF) was the fastest at 1.1 min/subtask but cost $4.51/subtask, while Gemini 3.8 Flash (H) was the cheapest at $0.38/subtask but slower at 4.18 min/subtask.
• Extrapolating current trends suggests LMs could compete with human teleoperators (25-40 seconds per task) by mid-2027.