Arena AI announced a new routing system called Max. The router decides which underlying model should handle each user request, based on how hard the task is and how much it will cost. In a live test Max topped the Agent Arena leaderboard with a 14.7% score boost and an average cost of $3.27 per task.
How Max works
The system looks at more than 2.5 million real‑world sessions to learn each model’s strengths. For code work it sends 67 % of sessions to Astra (Medium). For chat it splits the load between Astra (Max) (49 %) and Fable 5.1 (Max) (42 %). For general work tasks the split is 49 % to Fable 5.1 (Max) and 43 % to Astra (Medium). By matching tasks to the right model, Max moves the performance‑cost “Pareto frontier” upward, meaning better results for the same or lower price.
Early results and availability
In the same test Max beat the previous top public model, Fable 5.1 (Max), by 0.3 % while costing 37.5 % less. A prior experiment with Fable 5 and Kimi K3 showed a 10.7 % performance gain at a median cost of $1.06 per task, a 54 % saving. To celebrate the launch Arena AI is offering free access to the most capable Max variant (Fable 5.1 + Astra) in agent mode for a limited time. Users can also apply for API access via a signup form.
Why it matters
For developers and businesses using AI agents, Max can lower the price of each task while keeping or improving quality. That means larger projects can stay within budget, and smaller teams can experiment with advanced AI without paying premium rates.
