Multimodal Max

Max demonstrates strong performance on text, improving the time-to-first-token by more than 9 seconds compared to the next best model. The routing distribution is diverse, suggesting Max’s ability to leverage the various strengths of different models.

In Search, we find similar results, with Max able to achieve top performance on the leaderboard. The routing distribution is more concentrated on fewer models which are both strong in performance and in latency.

In Vision Arena, Max outperforms the best model at the time by 3 points while providing more than a 20 second speedup. The routing distribution shows Max strongly relying on gpt-5.2-chat-latest for 62% of battle prompts, but the 38% of prompts routed elsewhere netted Max a 12 point gain in strength.

In Code Arena, Max once again beats out its routing choices in terms of performance. Here we focused on making Max faster as measured by end-to-end latency since the output is only presented to the user upon full completion. Max leans more on claude-opus-4-5 variants than might be expected, largely due to gains in e2e latency.

Max in the Text-to-Image modality had an extremely strong performance, outperforming the top models in its routing set both on model strength and latency. The routing distribution leaned towards gemini-3.1-flash-image-preview, but the remaining routing choices were diversely spread through multiple models.

In the Single-Image Edit Arena, Max performed well, providing a faster but still strong alternative to gpt-image-2 (medium). Because the strength-latency tradeoff was configured to have a heavier emphasis on strength, Max still heavily relied on the top model. Interestingly, the routing distribution for Max in Image Edit Arena largely only consisted of other models on the pareto frontier, showing Max is able to identify models with the strongest latency/performance tradeoff.

Finally, on Multi-Image Edit Arena, Max also landed as a faster but still strong alternative to gpt-image-2 (medium). In this case, the strength-latency tradeoff was more heavily weighted towards speed, giving Max a 22 second speedup over gpt-image-2 (medium).

With these expanded capabilities, Max can be used for more diverse tasks. Whether you want to generate a graphic, interpret a chart, or make a live website, Max can handle it. Go give the new and improved Max a try!