← Back to feed

Arena.ai logged six model variants across three leaderboards on September 25 — seven Arena admissions in total, since GPT-6 Sol (Max) entered both Agent Arena and Text Arena simultaneously — the largest single-day batch since it launched competitive agent evaluation. The additions put GPT-6 Sol (Max) into head-to-head agent comparison for the first time, open a new front for StepFun in code evaluation, and give Text Arena its most crowded frontier cohort to date.

Sol Max Enters Agent Arena

GPT-6 Sol (Max) was added to Agent Arena on September 25 and simultaneously to Text Arena — the first time OpenAI’s Sol model has been submitted to agent-specific human evaluation. Agent Arena requires models to complete multi-step tasks involving tool use, code execution, and sequential reasoning judged by human raters, rather than the open-ended chat preferences measured in Text Arena.

Sol and Luna launched together on September 22 at a 50% price reduction from their GPT-5.6 predecessors. Sol (Max) is priced at $2 per million input tokens. Its most recent auxiliary benchmark data shows Terminal-Bench 4.0 at 43%, up 6 points from GPT-5.6 Sol max’s 37%. Claude Fable 5.1 sits at 57.9% on the same track. Agent Arena results will be the first human-preference signal for how Sol (Max) performs on agent tasks at that price point.

Claude Opus 5.5 high is also newly admitted to Text Arena as of September 25. Claude Opus 5.5 launched on September 22 at $4 per million input tokens with adaptive thinking enabled. Its LiveBench Agentic Coding score of 71.7 is the highest for any closed frontier model in the September run. The two models — Sol Max at $2 and Opus 5.5 high at $4 — will accumulate Arena votes in parallel, producing a head-to-head cost-adjusted comparison that no benchmark to date has provided.

Text Arena Gets Four More

Alongside Sol Max and Opus 5.5 high, Text Arena admitted three additional model variants on September 25: deepseek-v4.1-flash-max, grok-4.7-xhigh, and gpt-6-luna-max.

DeepSeek V4.1 Flash is already the top-ranked model on LiveBench Agentic Coding at 77.3, ahead of every closed frontier model. Its max tier variant entering Text Arena adds open-weight competition to a leaderboard that has historically been dominated by closed APIs. Grok-4.7-xhigh is xAI’s top reasoning mode for the 4.7 series, previously measured at 81k tokens of average output per task under extended reasoning. GPT-6 Luna (Max) is the smaller and faster sibling to Sol, priced at $0.10 per million input tokens.

Five distinct model variants entering a single leaderboard in one day is unusual. Arena ELO requires thousands of human votes to stabilize, and early votes carry higher variance than votes after rank stabilization. Entering a cluster of models simultaneously spreads early vote allocation across a larger set, slowing individual rank stabilization for each but establishing relative ordering between the cluster members faster than staggered entry would.

StepFun in Code Arena WebDev

Step-5-preview-high from StepFun was added to Code Arena WebDev on September 25. This is a separate variant from step-5-preview, which entered Code Arena WebDev earlier this month. The “-high” designation typically indicates a higher reasoning budget or effort tier rather than a different underlying model.

StepFun, a Chinese AI startup led by former Microsoft researchers, has not yet released SWE-bench Verified scores for the step-5 family through independent evaluators. Code Arena WebDev generates ranked human preference data for web development tasks — a narrower but more applied test than raw coding benchmarks. The track has over 130 model entries, currently topped by Claude Opus 5.5 Max.

Volume vs. Signal

The September 8-11 window saw three models enter across three Arena tracks in four days — GPT-6 Astra, Muse Spark 1.3, and DeepSeek V4.1 Flash. September 25 matched that track count in one day with six model variants.

The acceleration reflects a structural shift in how labs approach evaluation. Arena admission used to follow release by weeks or months. The September pattern — Sol and Opus 5.5 enter Arena within three days of launch — compresses that cycle to near-zero. Labs now appear to treat Arena admission as part of the launch sequence rather than a post-launch milestone.

Results from the September 25 cohort will require 30-60 days to reach statistical stability. The most consequential matchup will be Sol (Max) in Agent Arena versus Opus 5.5 in Text Arena — two models that launched on the same day at a 2:1 price ratio. Both appeared on LiveBench and the Artificial Analysis Intelligence Index at launch, but Arena evaluation applies human raters directly to task completion quality, where benchmark scores don’t always predict rank.