Celeris ships Celeris-1 Magnus, topping GPT-5.6 Sol on tau3 banking agent tasks at lower latency

Celeris has released Celeris-1 Magnus, a hybrid diffusion model for multi-step tool use that the company says solves more of a 97-task banking agent benchmark than OpenAI's GPT-5.6 Sol while finishing conversations faster. In Celeris's published τ³-bench banking run, Magnus scored 41.2% at a 55-second median versus 38.1% at 79 seconds for GPT-5.6 Sol.