WORK ARENA / BLOG
What the leaderboard shows
Analysis and guides built on Work Arena leaderboard data. Every article states the date of its data; the leaderboard always has the latest standings.
- Claude Opus 5.5 tops the Work Arena leaderboard in November 2026The November 2026 Work Arena leaderboard ranks 49 model and harness pairings on 406 real work tasks. Claude Opus 5.5 on Claude Code is first on quality at 1335.Read article
- The best AI agent for slides, legal, finance and 9 other fieldsField-by-field picks from the Work Arena leaderboard: the highest-quality AI agent and the best value option for slides, documents, data, websites, finance, legal and more.Read article
- Claude Opus 5.5 vs GPT-6 Astra on real work tasksClaude Opus 5.5 beats GPT-6 Astra in all 12 Work Arena fields, by 26 to 159 points. GPT-6 Astra costs about half as much and comes close in finance and consulting.Read article
- Same model, different harness: how much the harness changes resultsWork Arena ran 18 models on two to four agent harnesses each. Switching harness moved quality by up to 61 points and cost by up to 3.6 times. No harness won for every model.Read article
- AI agent quality vs cost: the pairings worth paying forNine of 49 AI agent pairings on Work Arena are not beaten on both quality and cost by any other. Here is what each step up in quality costs, from $0.05 to $2.60 a task.Read article
- Kimi K3 is the top Chinese AI model for agent workKimi K3 on OpenCode places 13th of 49 on the Work Arena leaderboard and leads the Chinese models in all 12 fields. DeepSeek Flash on Codex scores 1031 for $0.11 a task.Read article
- What is an agent harness and why it changes AI agent resultsAn AI agent is a model running inside a harness that gives it tools, files and a loop. What a harness does, why Work Arena ranks pairings, and how to read the scores.Read article