Arena Launches Agentic Router Max: Ties Flagship Performance at 37% Lower Cost

Arena has released Max, a new-generation agentic model router built for Agent Arena, claiming to match or even beat frontier-model performance at a much lower cost.

Confirmed

• Max was trained on over 2.5 million real Agent Arena sessions and dynamically picks the best model based on task complexity, cost, and type.
• It ranked #1 in live Agent Arena tests with a net improvement score of +14.7%. Blogger @infwinston added that its median per-task cost is $3.27, roughly 37% lower than the then-top public model Fable 5.1 (Max) (the original post said "higher by 0.3…" but the text was truncated).
• Arena's published Pareto frontier chart shows that by dynamically routing between Claude Fable 5.1 and GPT 6 Astra, Max outperforms either single model across Code, Chat, and Work task categories.
• Routing distribution: 67% of coding sessions go to the cheaper Astra (Medium); chat sessions favor Astra (Max) at 49% and Fable 5.1 (Max) at 42%; work sessions are spread mainly across combinations not fully shown in the post.

Why it matters

Max demonstrates the potential of the "dynamic per-task routing" approach: you can approach or exceed frontier-level performance without paying for the priciest model on every request, offering a quantifiable cost-control blueprint for agent applications.

2026-10-14 ~ 2026-10-14 · 6 related posts

• Arena's agentic router Max matches frontier performance at 37% lower cost — arena · 2026-10-14
• [source] Arena's Max router shifts the Pareto frontier across Code, Chat, and Work — arena · 2026-10-14
• Arena details Max routing: 67% of code sessions go to cheaper Astra (Medium) — arena · 2026-10-14

3 near-duplicate retellings: arena · infwinston · arena