← Back to feed

Arena added new model entries across five leaderboard categories on October 2 — a single-day update covering image generation, text, vision, and agent rankings for models from xAI, Anthropic, and OpenAI.

The Additions

grok-imagine-image-2.0 (canvas) joined Arena’s Text-to-Image and Image Edit leaderboards. The canvas variant is distinct from the original Grok Imagine Image 2.0 that entered Arena in August at second place in both image categories, behind GPT-Image-2. xAI’s canvas designation refers to image-layer compositing and in-context editing capability — a different capability profile from the original’s text-to-image generation.

claude-sonnet-5.5-xhigh entered the Text Arena and Vision Arena leaderboards, marking the first appearance of any Sonnet 5.5 variant in these two categories. Sonnet 5.5 launched September 28 at $2 per million input tokens and $10 per million output tokens. The xhigh effort tier is the second-highest compute mode — max sits above it — and the Arena entry will generate the first open human-preference ELO scores for Sonnet 5.5 in text and vision contexts.

Claude Sonnet 5.5 (Max) and GPT-6.1 Sol (Max) were added to Agent Arena on the same date, creating a direct head-to-head on agentic tasks. That addition was reported separately.

Context: Sonnet vs. Opus on Quality

Sonnet 5.5 sits two points below Opus 5.5 on Artificial Analysis’s Intelligence Index in their respective max-effort modes — Opus 5.5 Max scores 58, Sonnet 5.5 Max scores 56. At half of Opus 5.5’s output price ($10 vs. $20 per million output tokens), the gap is smaller than the price differential suggests. The Text Arena and Vision Arena ELO data for xhigh will reveal whether human raters are sensitive to that two-point gap in open-ended tasks, or whether preference converges when both models use extended thinking.

The Capability-Mode Expansion Pattern

The simultaneous additions across five categories in one day reflect how the evaluation landscape is expanding. xAI’s canvas variant and Anthropic’s xhigh setting are not new models — they are operating modes of existing models. Arena is increasingly treating these as distinct leaderboard entrants, recognizing that a model’s ELO under maximum extended thinking may differ meaningfully from the same model at a lower effort level.

Three separate xAI image entries now compete on Arena’s image leaderboards. Multiple Sonnet 5.5 effort variants are distributed across different Arena categories. The combinatorial growth in what gets evaluated is outpacing the growth in underlying model count.