Arena's agentic Max router, trained on 2.5M Agent Arena sessions, tops Agent Arena at a $3.27 median task cost — 37.5% below Fable 5.1 (Max)

Arena (formerly LMArena) extended its Max router from chat and multimodal tasks to agents. Long-horizon agent sessions are expensive, and Arena's testing of 60+ agentic models over four months showed sharp cost/performance differences across code, chat and work tasks. The new Max is trained on 2.5M+ real Agent Arena sessions and routes each request by task complexity, type and cost — in this release between Fable 5.1 (Max), Astra (Max) and Astra (Medium). It learns to spot costly-looking coding sessions the cheaper Astra (Medium) will likely solve: 67% of code sessions go there, while chat splits 49% Astra (Max) / 42% Fable 5.1 (Max) and work splits 49% Fable 5.1 (Max) / 43% Astra (Medium).

In a live test Max ranked #1 in Agent Arena with a +14.7% score at a $3.27 median cost per task, edging the top public model, Fable 5.1 (Max), by 0.3% at 37.5% lower cost. An earlier Fable 5 + Kimi K3 combination reached +10.7% net at a $1.06 median cost, 54% cheaper than Fable 5 (High). These are Arena's own measurements, and the router was trained on data collected before Sept 17, so it does not include Opus 5.5.

Max is free to try for a limited time by selecting "Max" in Agent Mode; API access requires a signup form. For teams building their own agents, the takeaway is tiered routing: send most coding steps to a mid-tier model and escalate only when needed.