Arena's Max router launches with Fable 5 and Kimi K3 at lower cost than single models
Overview
Arena announces Max, an agentic version of its model router that picks a model for each task, trained on 2.5 million sessions from Agent Arena.
With Fable 5 and Kimi K3, Max achieves a +10.7% net improvement at a median cost of $1.06 per task, which Arena says outperforms Fable 5 (High) by 0.6 percentage points while cutting cost 54%.
In a separate post, Arena says Max reached a +14.7% net improvement score at a median cost of $3.27 per task, matching Claude Fable 5.1 (Max) at 37% lower cost. Arena says it routes between Fable 5.1 and GPT 6 Astra and shifts the performance-cost frontier in its Code, Chat, and Work categories. Arena offers Max free for a limited time, and says the routing method works across any model combination.
Written by AI from the articles below · updated Oct 13, 4:34 PM ET
Check the sources:
Developments
4 developments
- Oct 13, 4:29 PM ET · 1 articleArena announces Max router achieving +10.7% net improvement at lower cost with Fable 5 and Kimi K3Arena: Arena's Max router beats Fable 5 (High) at 54% lower cost
- Oct 13, 4:17 PM ET · 1 articleArena reports Max routing sessions across Astra and Fable 5.1 by task typeArena: Max routes coding, chat and work tasks to different models
- Oct 13, 4:17 PM ET · 1 articleArena says its Max router shifts the cost-performance frontier in Agent Arena across Code, Chat, and WorkArena: Arena's Max routing shifts the performance-cost frontier in three task categories
- Oct 13, 4:17 PM ET · 1 articleArena launches agentic Max model router at 37% lower costArena: Arena launches agentic Max router, matching Claude Fable 5.1 at 37% lower cost
Article timeline
The articles in this story. Times are ET.
Arena.ai@arenaOfficialArena's Max router beats Fable 5 (High) at 54% lower costAIArena says its Max routing method, paired with Fable 5 and Kimi K3, achieves a +10.7% net improvement at a median cost of $1.06 per task. That outperforms Fable 5 (High) by 0.6 percentage points while cutting cost 54%. Arena says the same approach can work across other model combinations.

Arena.ai@arenaOfficialArena's Max routing shifts the performance-cost frontier in three task categoriesAIArena says its Max system moves the performance-cost Pareto frontier in Agent Arena's Code, Chat, and Work categories. By routing between Fable 5.1 and GPT 6 Astra, Max delivers stronger results at lower cost than any single frontier model.

Arena.ai@arenaOfficialMax routes coding, chat and work tasks to different modelsAIArena says Max routes 67% of code sessions to Astra (Medium), while chat favors Astra (Max) at 49% and Fable 5.1 (Max) at 42%. Work sessions split mostly between Fable 5.1 (Max) at 49% and Astra (Medium) at 43%. Max also learns which costly coding sessions Astra (Medium) can likely solve, routing those to it to save money.

Arena.ai@arenaOfficialArena launches agentic Max router, matching Claude Fable 5.1 at 37% lower costAIArena introduces Max, an agentic version of its model router that selects the best model for each task. Trained on 2.5M real-world sessions from Agent Arena, Max achieves a +14.7% net improvement score at a median cost of $3.27 per task. Arena says this matches Claude Fable 5.1 (Max) performance at 37% lower cost, and offers Max free for a limited time.

Heat trend
Not enough continuous observations to show a trend yet.