Arena 榜单更新：Fable 5.2 登顶 Agent 榜，Gemini 4 Hydrogen 主打速度
This Week in the Arena: Fable 5.2 takes #1, and Gemini 4 Hydrogen brings the speed. - Claude Fable ...
这期 Arena 周报信息量很密：Fable 5.2 在 7.8k 真实会话里涨了 30.5%，Gemini 4 Hydrogen 一步只要 7.3 秒。选模型做智能体或写代码前，先看看这份榜单再定，ERNIE 5.5 Flash 和 Inkling 2 Small 的开源成绩尤其值得对照。
Agent Arena 最新周榜显示，Anthropic 的 Claude Fable 5.2 (Max) 以 +30.5% 的净提升登顶第一，数据基于 7.8k 条真实任务会话。Google DeepMind 的 Gemini 4 Hydrogen (High) 同期上线，成为 Agent Arena 前 10 中最快的模型，单步典型耗时 7.3 秒。
开源阵营方面动作密集：thinkymachines 的 Inkling 2 Small 开放权重，以 1647 分成为 Code Arena: WebDev 榜第一的美国开源模型；百度 ERNIE 5.5 Flash Preview 目前在 WebDev 自动评测中约 1668 分，正逼近开源第二，配备 1M token 上下文窗口，Apache 2.0 开源权重即将放出。Meta 最小的模型 Muse Ember (Max) 首次进入 WebDev 榜，排第 37，得分 1609，每任务中位成本 $0.03。图像编辑榜上，MiniMax H3-Image 以 1365 分居开源第一（总榜第 11），与 Nano Banana 2、Nano Banana Pro 打平。
对开发者而言，这轮更新意味着开源模型在代码与图像编辑两类任务上正在缩小与闭源头部模型的差距，而闭源侧则在推理速度和智能体任务表现上继续拉开身位。
This Week in the Arena: Fable 5.2 takes #1, and Gemini 4 Hydrogen brings the speed. - Claude Fable ...
This Week in the Arena: Fable 5.2 takes #1 , and Gemini 4 Hydrogen brings the speed. - Claude Fable 5.2 (Max) by @AnthropicAI takes #1 in Agent Arena with +30.5% net improvement across 7.8k real-world sessions as of today. - Gemini 4 Hydrogen (High) by @GoogleDeepMind launched as the fastest model in the Agent Arena top 10, with a typical step of 7.3 seconds. - Muse Ember (Max) by @AIatMeta , Meta’s smallest model, debuted in Code Arena: WebDev. It’s #37 with 1609 points today, at a $0.03 median cost per task. - Inkling 2 Small by @thinkymachines is now open weights and the #1 US open model in Code Arena: WebDev at 1647 points. - MiniMax H3-Image by @MiniMax_AI is the #1 open model in Multi-Image Edit at 1365 ( #11 overall), on par with Nano Banana 2 and Nano Banana Pro. - ERNIE 5.5 Flash Preview by @Baidu_Inc is trending toward #2 open model in Code Arena: WebDev at 1668 points (AutoEval), with a 1M-token context window. Open weights under Apache 2.0 are coming soon. youtube.com/live/FD3zZ5X0k…📊 3 ⚡
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Claude Fable, Gemini 4, ERNIE, MiniMax.
Source attribution: curated from lmarena.ai (@lmarena_ai). Cross-source corroboration: 11 independent sources have covered this development (including TestingCatalog, 宝玉, 岚叔), confirming industry-wide signal strength.
