LM Arena's Max Router Tops Agent Benchmarks While Cutting Costs 37.5%

Arena has released an upgraded version of Max for multi-turn agent workflows. Earlier versions handled chat and image tasks; the new router selects models for longer sessions that may involve code execution, tool calls, retries, and accumulating API costs. Max predicts which model can complete each task based on its type, complexity, and expected cost.

The router targets the performance-cost Pareto frontier, meaning configurations where quality cannot improve without higher spending and costs cannot fall without lower quality. In practice, Max tries to send routine work to less expensive models while reserving premium models for tasks that need them.

Live results: $3.27 per task

In a live Agent Arena test, Max ranked first with Arena’s reported score of +14.7% and a median cost of $3.27 per task. It outscored the leading public model, Fable 5.1 (Max), by 0.3 percentage points while costing 37.5% less.

The cost reduction puts Max at roughly five-eighths of the comparator’s bill while maintaining slightly higher performance in Arena’s evaluation. Because $3.27 is a median, half of the measured tasks cost more and half cost less; Arena’s published figures do not show average or tail costs.

Three signals shape each route

Max learned its routing policy from more than 2.5 million Agent Arena sessions. It uses task complexity, task category, and expected cost to choose among the available models, producing different routing patterns for code, chat, and workplace tasks.

Arena’s summary lists the leading destinations rather than every route, so the percentages do not total 100% in each category. The distributions show that Max does not apply one fixed model hierarchy across workloads.

The clearest cost-saving pattern appears in coding sessions. Max identifies potentially expensive jobs that Astra (Medium) is likely to complete and sends them to the cheaper model, reducing the cost of long runs without assigning every coding task to a premium tier.

Model pools can be replaced

Arena describes Max as model-agnostic because the routing method can be trained for different pools. In an earlier test pairing Fable 5 with Kimi K3, Max reported a +10.7% net improvement and a median cost of $1.06 per task. It exceeded Fable 5 (High) by 0.6 percentage points while cutting costs by 54%.

These first-party tests show that Arena has applied the method to two model combinations. Each new pool still requires current training data, evaluation, and deployment work before the router can select its models.

Long sessions amplify mistakes

Agent routing carries higher stakes than single-turn chat routing because sessions can span many model calls and external tools. A poor early decision may trigger retries, unnecessary tool use, or a failed workflow after substantial compute has already been consumed.

Arena previously released the original Max and a multimodal router for chat and image tasks. The company says it has tested more than 60 agentic models on its platform during the past four months, giving it a broad pool of session data for training and comparison. The reported model count alone does not establish performance on workloads outside Agent Arena.

Model routing has become a distinct infrastructure category because task requirements vary more than model pricing suggests. Routine jobs may complete successfully on mid-tier models, while difficult sessions benefit from premium capacity. OpenRouter’s auto mode and Martian pursue similar per-task selection strategies.

Access now, API later

Arena currently offers free, limited-window access to the top Max variant in Agent Mode. That version routes between Fable 5.1 and Astra. The company has placed an API on its roadmap and offers a waitlist, but it has not announced an API release date or production pricing.

Production tests need local data

1. Evaluation fit: Max’s gains come from Arena’s own framework, which emphasizes chat, code, and work tasks. A different workload mix may produce different routes, costs, and completion rates.
2. Training freshness: The Fable 5.1 and Astra router was trained on data collected before September 17, prior to the release of Opus 5.5. Later models require a refreshed pool and updated training before Max can select them.
3. Latency and reliability: Routing introduces another decision and network step. Latency-sensitive agents should measure router overhead, timeout behavior, retries, and fallback handling.
4. Missing distribution data: The published results provide medians and aggregate scores but omit p95 cost, p95 latency, category-level success rates, and failure distributions.

A production evaluation should replay representative traces against both Max and a fixed-model baseline. Completion rate, total model and tool cost, p95 latency, retries, and unrecovered failures will show whether Arena’s reported 37.5% saving carries over to the target application.