Sonnet 5.5 and GPT-6.1 Sol debuts reshape the Agent Arena leaderboard
This week's LMArena update brought notable shifts to Agent Arena: Anthropic's Claude Sonnet 5.5 (Max) debuted at #3 with a +12.5% net improvement score, while OpenAI's GPT-6.1 Sol (Max) took #5 with +11.23% while sharply cutting per-task cost, together reshaping the performance-cost Pareto frontier.
Confirmed
- Agent Arena is based on real Agent Mode tasks, with 2.15M+ sessions across 51 models; real-time leaderboard and Pareto data are published by Arena officially.
- Claude Sonnet 5.5 (Max) debuted at #3 with +12.5% net improvement and $2.74 median cost per task; 8.1 percentage points above Claude Sonnet 5 (High) at #13 (+4.4%), but its per-task cost is 73% higher than Opus.
- GPT-6.1 Sol (Max), added after OpenAI DevDay, ranks #5 with +11.23% net improvement at $0.56 median cost per task, 39% cheaper than GPT-6 Sol (real-time data in another post: $0.57 per task).
- The Pareto frontier also includes Claude Fable 5.1 (Max) at +14.31% and $4.62 per task.
- On the text side, Gemini 4 Argon topped Text Arena; another post notes Sonnet 5.5 approaching GPT-6 at roughly a 20% price gap, with four new models reshaping the Pareto frontiers of Text, Code and Agent Arenas.
Why it matters
- Sonnet 5.5 enters the top tier of performance, but its $2.74 per-task cost makes its price-performance notably poor, contrasting with its near-parity with GPT-6 on price in Code Arena—users must weigh performance against cost carefully.
- GPT-6.1 Sol reaching the top five at well under the previous cost shows frontier agent capability prices are falling fast, and the reshaped Pareto frontier may change model selection decisions for enterprises and developers.
2026-10-03 ~ 2026-10-03 · 5 related posts
Primary sources
- Claude Sonnet 5.5 debuts at #3 in Agent Arena, costing 73% more per task than Opus 5.5 — arena ·
- GPT-6.1 Sol hits #5 on Agent Arena, costs 39% less than GPT-6 Sol while scoring higher — arena ·
- Agent Arena leaderboard: GPT-6.1 Sol joins Pareto frontier at $0.57/task, DeepSeek cheapest open model — arena ·
- Arena Weekly: Gemini 4 Argon Tops Text Arena, Sonnet 5.5 Within 2 Points of GPT-6 at 80% Less Cost — arena · 2026-10-03
- [source] Claude Sonnet 5.5 debuts at #3 in Agent Arena, costing 73% more per task than Opus 5.5 — arena · 2026-10-03
- Claude Sonnet 5.5 Lands #3 on Agent Arena at $2.74/Task, Just Off the Pareto Frontier — arena · 2026-10-03
- [source] GPT-6.1 Sol hits #5 on Agent Arena, costs 39% less than GPT-6 Sol while scoring higher — arena · 2026-10-03
- [source] Agent Arena leaderboard: GPT-6.1 Sol joins Pareto frontier at $0.57/task, DeepSeek cheapest open model — arena · 2026-10-03