Six Frontier Models Enter Vision Arena on October 9 — All Five Major Labs Now Represented
Arena.ai’s October 9 Vision leaderboard update is the most significant single-day expansion the visual evaluation track has seen. Six models from five labs entered simultaneously, giving Vision Arena the same cross-lab coverage that the Text and Agent arenas have had for months.
The Six Additions
| Model | Lab | Effort Tier |
|---|---|---|
| claude-opus-5.5-high | Anthropic | High |
| muse-spark-1.3 | Meta | xHigh |
| gpt-6-sol-max | OpenAI | Max |
| deepseek-v4.1-flash-max | DeepSeek | Max |
| grok-4.7-xhigh | xAI | xHigh |
| gpt-6-luna-max | OpenAI | Max |
All six additions land on the Vision Arena leaderboard, which evaluates visual understanding and reasoning through blind pairwise human preference votes — distinct from the Text-to-Image and Image Edit leaderboards that track generation quality.
What Each Model Brings
Claude Opus 5.5 is Anthropic’s current top model, holding the number one position on Artificial Analysis’s Intelligence Index as of late September. Its text benchmark profile is strong: LiveBench overall 83.2%, OSWorld 2.1 partial 81.8%, HLE with tools 67.7%. Its Vision Arena debut will give the first external read on how its visual reasoning stacks up against the rest of the field.
Muse Spark 1.3 is Meta’s frontier offering, scoring 81.6% overall on LiveBench September 2026, tied with GPT-6.1 Sol. The xHigh effort tier indicates extended compute at inference. Muse Spark entered the text leaderboard in September and had not yet been evaluated in Vision Arena until October 9.
GPT-6 Sol is OpenAI’s reasoning-optimized model priced at $2 per million input tokens and $10 per million output. It had already appeared in Agent Arena in late September; the October 9 Vision Arena addition extends the head-to-head to the visual modality. DeepSWE v1.1 at max effort: 68.8%.
DeepSeek V4.1 Flash is the only open-weight model in the October 9 cohort. Its LiveBench Agentic Coding score of 77.3% leads the September leaderboard, more than 6 points above Claude Opus 5.5’s 71.7%. Its Vision Arena debut is the first structured external visual evaluation for the model.
Grok 4.7 is xAI’s most recent model, launched September 21. It scored 93.15% on VulcanBench Frontier v4 in independent testing from October 1–3 and 71.0% on DeepSWE v1.1 at high effort. Pricing: $2 per million input, $6 per million output.
GPT-6 Luna is OpenAI’s cost-optimized counterpart to Sol, priced at $0.10 per million input tokens and $0.50 per million output — the most accessible model in the October 9 cohort by an order of magnitude. Terminal-Bench 4.0 at 13%. Its Vision Arena entry positions Luna as the efficiency reference point for visual tasks.
Context
Before October 9, the Vision Arena’s frontier coverage was incomplete. Claude Sonnet 5.5 xHigh was added to Vision Arena on October 2 alongside its Text Arena entry. Prior to that, coverage skewed toward text and agent leaderboards, where multi-lab head-to-heads had been running since earlier in Q3.
The October 9 batch closes the gap. With Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, Grok 4.7, Muse Spark 1.3, and DeepSeek V4.1 Flash all entering on the same day, Vision Arena now mirrors the cross-lab field already competing in text and agent tracks. Initial ELO scores will accumulate over the following days as Arena users run pairwise visual comparisons.
The DeepSeek V4.1 Flash inclusion is notable. It is the first time an open-weight model with frontier-tier agentic coding performance has been entered into Vision Arena alongside closed-API rivals at their most capable effort tiers. How it ranks on visual tasks relative to labs with proprietary training pipelines is an open question.