Arena adds agent routing as new models reshape its leaderboard

Arena has introduced an agent-focused version of its Max router alongside new Agent Arena entries from Google, Xiaomi, and Meta. Arena says Max can match the task performance of Claude Fable 5.1 (Max) while reducing median cost by 37%.

For developers running multi-step agents, the update turns model selection into a per-task infrastructure decision. Max chooses among backing models according to expected performance and cost, while Agent Arena measures the results across sessions that can include tool calls, code execution, retries, and follow-up turns.

Max picks a model per task

Arena trained Max on 2.5 million sessions that it describes as real-world agent interactions. The router analyzes each incoming task and selects the model with the strongest expected quality per dollar.

The percentage is Arena’s internal “net improvement” metric. The published rundown does not define its calculation or baseline, so the figure cannot be translated directly into an accuracy gain.

The disclosed route shares show that Max changes its preference by workload:

The published shares omit less frequently selected routes, which account for the remaining sessions. Arena also describes the design as model-agnostic, allowing it to update the available model pool as prices and capabilities change.

Routing gains must clear the overhead

A router creates value only when its model choices save more time or money than the routing layer consumes. Poor selection can increase token use, add latency, and send a task to a model that requires more repair turns.

In Arena’s recent evaluation of an unnamed competing router, comparable end-task performance cost 38% more than calling DeepSeek V4.1 Flash (Max) directly. Median request latency was 1.7 times the direct call’s latency.

Max’s training data covers multi-turn traces, giving the router signals from code execution, tool use, retries, and accumulated API costs. Those traces may help it identify tasks where a less expensive model can complete the full session successfully. Arena does not specify whether Max selects one model for an entire task or can switch models between turns.

Arena currently offers Max free in Agent Mode for a limited period, with no end date stated. API access requires joining a waitlist.

Argon climbs as Meta and Xiaomi add contenders

Google’s Gemini 4 Argon (High) leads several Agent Arena categories in the weekly update:

Steerability measures how well a model adjusts after a user corrects or redirects it. Confirmed Success tracks sessions where the requested outcome was verified, while the Chat score covers conversational tasks.

Meta’s Muse Wonder (Max) holds No. 7 overall with an Elo rating of 1501. Xiaomi’s mimo-v2.6-pro ranks No. 11 at 1491, making MiMo the highest-ranked open model according to Arena. Elo is a relative rating, so a model’s score depends on its results against other models in the pool.

Meta has also released Muse Spark Open, an 816-billion-parameter open-weight model. It is live in Agent Arena, but its score is still accumulating.

Pricing and snapshots blur Argon’s position

Arena’s rundown gives Gemini 4 Argon two overall positions: No. 5 in the weekly summary and No. 8 in a later comparison. The update provides no snapshot timestamps or methodological distinction that would reconcile those ranks.

Argon also matches GPT-6 Astra on Artificial Analysis’s overall index while trailing Claude Opus 5.5. Differences between that index and Agent Arena can reflect separate task sets, scoring systems, model settings, and evaluation dates.

Argon’s promotional API price is $2 per million input tokens and $10 per million output tokens through November 30. The rates then double to $4 for input and $20 for output. Any cost-performance comparison based on the introductory rates must be recalculated for workloads running after the increase.

Benchmark the full agent loop

Teams evaluating Max or Argon should measure complete sessions on representative production traces. A useful comparison should include:

• Task success and verified completion rates
• Total input and output tokens
• Tool calls, retries, and repair turns
• Median and tail latency
• Median and high-percentile session cost
• Cache behavior and routing overhead
• Results under Argon’s promotional and standard prices

Production adoption also requires checking which providers Max can route to, whether teams can restrict that pool, how failures are retried, and which data-handling policies apply. Arena’s reported figures make the router testable, but they leave those operational details and tail-cost measurements open.