With our original Max and Multimodal Max releases, we explored model routing across chat-based and multimodal tasks. This release continues our research into intelligent routing, tackling the challenge of balancing performance and cost across multi-turn, long-horizon tasks in Agent Arena.
Powered by 2.5+ million real-world Agent Arena sessions, Max routes users’ requests to the best models based on task complexity, cost and type of tasks. By accounting for each model’s strengths and performance-cost tradeoffs, Max achieves Pareto-frontier performance across task categories.
In the last 4 months, we’ve tested over 60 different agentic models on Arena, revealing distinct trade-offs between performance and cost, as well as the strengths and weaknesses of models across chat, code, and work tasks. Max matches each task with the model best suited to handle it, delivering a consistent, reliable experience across chat, work, and coding.
We recently deployed the agentic Max router in a live test, where it ranked #1 in Agent Arena with a score of +14.7% and a median cost of $3.27 per task over its testing period. Max outperformed the #1 public model at the time of testing, Fable 5.1 (Max), by 0.3% at 37.5% lower cost*. Below, we share further insights and a breakdown of Max’s routing distribution.
Max Router Breakdown
Max shifts the performance–cost Pareto frontier across code, chat, and work, with the largest gains in chat and work. The solid lines show the frontier with Max, compared with the dotted lines for public models alone. By routing between Fable 5.1, Astra medium and Astra max, Max delivers stronger results at a lower cost than directly selecting these models.
Max adapts its routing to each task: 67% of code sessions go to Astra (Medium), while chat favors Astra (Max) at 49% and Fable 5.1 (Max) at 42%. Work sessions mostly split between Fable 5.1 (Max) at 49% and Astra (Medium) at 43%. Max learns to identify potentially costly coding sessions that the less expensive Astra (Medium) is likely to solve successfully, selectively routing those sessions to maximize savings.
Max routes across any model combinations
We designed Max’s routing methodology to be extensible to different sets of models, allowing the router to stay on the frontier even as new models launch. In a previous test with Fable 5 and Kimi K3, Max achieves a +10.7% net improvement at a median cost of $1.06 per task, outperforming Fable 5 (High) by 0.6 percentage points while reducing cost by 54%. These results demonstrate how the same approach can improve the performance-cost frontier across different model combinations.
Try Max in Agent Mode today
To celebrate its launch, we’re offering direct access to our most intelligent variant of Max (Fable 5.1 + Astra) in agent mode for a limited time. Bring your most ambitious project and see what Max can do!
Select “Max” in Agent Mode to get started.
Interested in accessing Max via API? Fill out the signup form here.
*Max Router (Fable 5.1 + Astra) was trained on data collected before September 17, prior to the release of Opus 5.5.
