token efficiency = cost efficiency.
why pay more when you can pay less
agent-device delivers more finished work for every dollar.
Benchmarked on @swmansionβs AppControlBench with GPT-5.4-mini and Haiku 4.5:
β ~4Γ more completed tasks/$
β 94.2β98.3% task completion
β 2% faster with GPT-5.4-mini; 26% slower with Haiku 4.5 than alternative
