← Back to feed

Arena (the organization formerly known as LMSYS behind Chatbot Arena) has shipped the multimodal expansion of Max, its latency-controlled model router built on 5 million+ community battle votes. Starting this week, Max is the default option for all modalities in Arena’s direct chat interface.

Max is not a model. It is a router: for each user prompt, it selects the highest-capability model for that specific request from a routing set of frontier models, subject to a latency constraint. The original Max, released in February 2026, demonstrated that routing outperforms individual models in text — the multimodal expansion tests whether that advantage holds across harder modalities.

The Benchmark Results

Max holds Pareto frontier performance across all seven modalities it covers:

Text: Top performance on the leaderboard. Routing distribution is spread across several models, with no single dominant choice.

Search: Similar concentration profile. Max achieves top performance with fewer models driving the routing decisions.

Vision: Max outperforms the best individual model by 3 points while providing a 20+ second latency advantage. The routing distributes 62% to gpt-5.2-chat-latest with the remaining 38% routing elsewhere gaining 12 additional ELO points in aggregate.

Text-to-Image: Extremely strong. Max outperforms top individual models on both strength and latency. Routing leans toward gemini-3.1-flash-image-preview with diversified allocations elsewhere.

Code (Front-End): Max again beats its routing set on performance. Latency emphasis here was configured for end-to-end speed since code output is only shown on completion. Notably heavy reliance on claude-opus-4-5 variants for latency gains.

Single-Image Edit: Second place, not first — but provides a large latency benefit over the top model (gpt-image-2 medium). Strength-latency tradeoff was configured toward strength.

Multi-Image Edit: Also second. Heavier speed weighting gives a 22-second advantage over gpt-image-2 medium.

Why This Matters

Most leaderboard comparisons assume you have to pick one model. Arena’s data — millions of real user battles — shows that no single model is uniformly best across all prompt types even within a single modality. The routing insight is that the correct model for each prompt is knowable before inference completes.

The meta-finding is about the evaluation org building production product. Arena is simultaneously:

• Running the most credible AI evaluation leaderboard
• Publishing the model routing weights that beat any individual entrant
• Deploying those weights as the default interface for users

This creates a structural position: Arena accumulates signal that no individual lab can match because the evaluation is multi-lab by design, and that signal compounds into routing intelligence that outperforms any single lab’s model.

Technical Construction

Max is updated periodically as new frontier models become available for routing. The benchmarks represent a point-in-time snapshot of which models were publicly routable when Max was last trained. The routing distribution for each modality reflects Max’s learned policy under latency constraints — models that offer the best strength-per-millisecond at each prompt type receive higher allocation.

The published results include the routing distributions (which models get what share of traffic) for each modality, which functions as an implicit capability benchmark: high routing allocation under a latency constraint is a strong signal of real-world value.

Front-end coding received dedicated sub-leaderboard analysis based on 250,000+ Code Arena prompts categorized into major task types. Results show material performance differences across task categories that aggregate leaderboards conceal.