13 posts tagged by Digg
Technology
Arena says its index draws on real-world agent traces, with results for over 20 frontier models at launch.
AI
Arena says its index tracks unauthorized actions, false attribution and deceptive completion across 90,000-plus agent sessions and 27 models.
AI
Arena said the Direct Mode test would start at 8 a.m. PT, with the model selectable from a dropdown menu. It said the model was already available in Battle and Agent modes.
AI
Arena says users can test GPT-6.1 in Battle Mode or Agent Mode. Votes will shape its evaluation, and Arena said scores would follow.
AI
Arena says the showcase traces model progress from Q4 2025 to September 2026. It features Claude, GPT, Gemini, Kimi and Qwen models, with scores for Claude Sonnet 5.5 and GPT-6.1 coming soon.
AI
Arena says it compared 34,580 model verdicts with human votes across 1,460 battles. On average, a judge chose its own answer 58% of the time, while people chose that same answer 34% of the time.
AI
Agent Arena's later update puts the model at +1.59% net improvement and a $0.05 median cost per task, reversing an earlier post that said it had missed the frontier.
AI
Arena says UC Berkeley PhD candidate Melissa Pan explored the “harness tax”: how the system around an AI model affects its cost and performance.
AI
Arena said users could select OpenAI’s GPT-6 Sol (Medium) in Direct Mode until 9 a.m. PT on September 29. It said the model would remain available in Battle and Agent Mode afterward.
AI
Black Forest Labs posted the updated model's Arena.ai leaderboard position after internal testing.
AI
OpenAI GPT-5.6 variants demonstrate gains on real-world agent tasks via community evaluations.
AI
Arena.ai ranks models on leaderboards using new factuality metric that blends human preferences with verified accuracy.
AI
Arena.ai ranks Kimi K3 Max first among open-weight models across agent evaluations.
