Arena released Max on October 12, 2026, as an agentic router for Agent Arena, saying it delivered frontier performance with 37% cost savings compared with top models. Built from 2.5+ million real-world Agent Arena sessions, Max selects models based on task complexity, cost and task type. Arena said the router achieved Pareto-frontier performance across task categories, balancing performance and cost.

Over the previous 4 months, Arena tested more than 60 agentic models across chat, code and work tasks, finding different performance-cost trade-offs and model strengths. In a live test, Max ranked #1 in Agent Arena with a +14.7% score and a median cost of $3.27 per task. It beat Fable 5.1 (Max), the top public model at the time of testing, by 0.3% while costing 37.5% less. Arena says the test used data collected before September 17, before Opus 5.5 was released.

Max’s routing differed by task: 67% of code sessions went to Astra (Medium); chat sessions went to Astra (Max) 49% of the time and Fable 5.1 (Max) 42%; work sessions went to Fable 5.1 (Max) 49% and Astra (Medium) 43%. Arena says Max identifies coding tasks that the less expensive Astra (Medium) can likely solve, then routes them there to save costs. The company says the router’s methodology can work with different model sets and keep performance on the frontier as models launch.

In an earlier test using Fable 5 and Kimi K3, Max achieved a +10.7% net improvement at a median cost of $1.06 per task. It outperformed Fable 5 (High) by 0.6 percentage points while cutting cost by 54%. Arena is offering limited-time access to Max, combining Fable 5.1 and Astra, through Agent Mode; people interested in API access can submit a signup form, according to the announcement. The announcement also describes Max as Arena’s newest version, building on its earlier Max and Multimodal Max releases.

Arena released Max on October 12, 2026, as an agentic router for Agent Arena, saying it delivered frontier performance with 37% cost savings compared with top models. Built from 2.5+ million real-world Agent Arena sessions, Max selects models based on task complexity, cost and task type. Arena said the router achieved Pareto-frontier performance across task categories, balancing performance and cost.

Over the previous 4 months, Arena tested more than 60 agentic models across chat, code and work tasks, finding different performance-cost trade-offs and model strengths. In a live test, Max ranked #1 in Agent Arena with a +14.7% score and a median cost of $3.27 per task. It beat Fable 5.1 (Max), the top public model at the time of testing, by 0.3% while costing 37.5% less. Arena says the test used data collected before September 17, before Opus 5.5 was released.

Max’s routing differed by task: 67% of code sessions went to Astra (Medium); chat sessions went to Astra (Max) 49% of the time and Fable 5.1 (Max) 42%; work sessions went to Fable 5.1 (Max) 49% and Astra (Medium) 43%. Arena says Max identifies coding tasks that the less expensive Astra (Medium) can likely solve, then routes them there to save costs. The company says the router’s methodology can work with different model sets and keep performance on the frontier as models launch.

In an earlier test using Fable 5 and Kimi K3, Max achieved a +10.7% net improvement at a median cost of $1.06 per task. It outperformed Fable 5 (High) by 0.6 percentage points while cutting cost by 54%. Arena is offering limited-time access to Max, combining Fable 5.1 and Astra, through Agent Mode; people interested in API access can submit a signup form, according to the announcement. The announcement also describes Max as Arena’s newest version, building on its earlier Max and Multimodal Max releases.