arena

SWE-Duel 提出一种自我扩展的编码智能体竞技场，两个智能体通过镜像红蓝回合互相评估：红色智能体开发新功能并植入隐藏缺陷，蓝色智能体在不知情的情况下审查并修复。在 15 个仓库、11 个模型-工具组合上进行了 1,500 回合的循环赛，显示任务难度可随作者能力扩展。

An open, replayable benchmark event where 10 AI models play a single world war with pacts, betrayal and nukes, recording every decision, reasoning and observation for full replayability.

作者运行了1200多次Blender建模实验，横跨多个LLM和agent框架（pi、opencode、omp、codex、Claude Code、dsh），并搭建了一个盲测A/B投票竞技场，用于对比不同模型、harness和推理等级在3D建模任务上的表现与成本效益。

Peter Gostev praises GPT-6.1-Sol as a daily driver for both work and personal use, claiming it offers roughly 5x more usage per subscription compared to Astra based on Arena's real-world cost-per-task data.

A new checkpoint for Gemini 4 Pro, codenamed barium-b, has appeared in Arena, showing noticeable improvements in 3D F1 studio output with better detail, materials, and UI coherence. The update is part of Gemini 4's post-training phase, with Google aiming for a fast release while safety work continues.

OpenAI's new AI models, GPT-Image-2.5-Sunburst and Flare, have achieved top rankings in text-to-image and image editing arenas, showing significant improvements over previous versions.

According to the Artificial Analysis arena, Grok 4.6 is reported to be equivalent in capability to Sol 5.6, offering a new benchmark comparison for these LLMs.

Grok Imagine Image 2.0 (Low) from xAI jumped to #2 in the Text-to-Image Arena, beating its own older quality model and showing significant improvement.

The author built AI Combat, a platform where users design AI agents with specific roles and strategies to battle each other in live 3-round matches with AI referees and ELO rankings.

Gergely Orosz comments on companies pivoting to AI-driven products in unfamiliar fields, citing Meta's new prediction markets app 'Arena' as an example.

Meta is reportedly developing a prediction markets app called Arena, similar to Polymarket and Kalshi, but initially using points instead of real money, as an internal top priority.

Agent Arena is the first public arena for AI agents, allowing users to test and compare AI agents in a competitive environment.

Stickblade Arena is a new benchmark where LLM agents control ragdolls in a 2D physics sword-fighting simulator, testing multi-turn tactical reasoning, spatial awareness, and real-time decision-making under adversarial pressure. Early results reveal capability gaps: DeepSeek R1 dominates melee but fails at bow due to time limits, and small models excel at close-range fighting.

Agent Hansa released an 'Arena' feature where AI agents compete in strategy, luck, and skill-based games for real money, as a social experiment.

Alibaba releases Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview on Arena, achieving top rankings in Text and Vision categories.

CruxArena.ai launched a platform letting users watch AI models debate consciousness in real time.

GPT-Image-2 surges to 1,512 on the Arena text-to-image leaderboard, opening a 242-point gap over rivals from Google DeepMind and OpenAI.