#评测/基准

Mercury 3 is live: the most capable diffusion LLM released to date. It outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running at over 1,000 tokens per second. More than 40x the size of Mercury 1, 3x more intelligent across standard benchmarks, and just as fast, with a 1M-token context window. Available today on the Inception API, @baseten, and @OpenRouter .

激动人心的消息：@deepseek_ai 的 DeepSeek-V4.1-Flash-20261017 (Max) 现已成为 Code Arena: WebDev 中排名第一的开源模型，以低 10 倍的成本超越 Kimi K3 (Max)！ 亮点： - 排名第一的开源模型，领先于 Kimi K3 (Max) - 以 1670 分位列总榜第 12，领先于 Claude Opus 5 (High)，与 GPT-6 Sol (Max) 相差约 10 分以内 - 以每百万输入/输出 token $0.3/$1.2 的价格推动帕累托前沿移动 - 相比上一版 DeepSeek-V4.1-Flash-20260910 提升 52 分 祝贺 @deepseek_ai 团队为开源生态又做出了一项强有力的贡献！

由 @AIatMeta 推出的 Muse Wonder (Max) 已进入 Code Arena：WebDev 以 1651 分位列第 15，Text Arena 以 1513 分位列第 2！ 在 Code Arena：WebDev 中，Muse Wonder (Max) 以 1651 分位列第 15。这一结果略低于以 1659 分位列第 12 的 Muse Spark 1.3 (Max)，与以 1655 分位列第 14 的 Kimi K3 (Max) 相差 4 分以内，并比以 1640 分位列第 16 的 Grok 4.7 (xHigh) 高 11 分。 此次发布在 Text Arena 中相较过去的 Muse 模型有显著提升。与以 1494 分位列第 11 的 Muse Spark 1.3 (Max) 相比，Muse Wonder (Max) 提升了 +19 分并上升 9 位至第 2。它落后于以 1525 分排名第 1 的 Gemini 4 Argon (High) 12 分。 祝贺 @AIatMeta 团队发布这一版本！

MiniMax M3.1 Flash：太空级性能 + 兔子级速度 🚀🐰 没错，Space Bunny——曾以 200T+ tokens 登顶全球排行榜——正是 M3.1 Flash 的早期版本。 💰重新定义智能体 AI 的性能经济学。M3.1 Flash 在 Cowork（Automation-Bench：62%；ALE：30%）、编程（DeepSWE：74%）和视觉推理（BabyVision：91%）等关键任务上达到前沿级智能水平，成本仅为 M3 的 1/3，速度提升 3 倍。 API 定价（统一费率）：每 1M 输入/输出 tokens 分别为 $0.15 / $0.60，缓存读取为 $0.015 / 1M tokens。 模型权重将于 10 月 23 日发布。 API：http://platform.minimax.io M Plan：https://www.minimax.io/m-plan MiniMax Code 桌面版：https://agent.minimax.io/download

OPENAI 🔥: 3 new model references have been spotted on the APIs recently. - aurora-red-alpha - aurora-green-alpha - aurora-blue-alpha > These model references were removed shortly after so it was a mistake rollout. > The chance that normal users will have a chance to test these exact models anytime soon is quite low (alpha). > Higher probability of a new Daybreak update announcement. This also doesn’t exclude the fact that other model releases might be planned. Not soon? 👀

🚀 DeepSeek-V4.1-Flash-1017 来了！ 与上一版本相比，基准测试表现更强、任务完成更快、输出更短——在性能与成本上实现了帕累托改进。现已在我们的网站、应用和 API 上线。 🛠️ 针对 @DeepSeekHarness、Tencent CodeBuddy、WorkBuddy 等进行了优化。升级到 DeepSeek Harness v0.2.1-rc.1 或更高版本，以获得新模型的最佳体验。 1/4

Typesafe's CEO @CompleteSkeptic does not trust public benchmarks. Neither do I, Diogo. So @p0 and I wrote a blog evaluating 8 decision models on real company training datasets we have personally audited and iterated on. For what it's worth though, Jev is still on top 👑 http://parallel.ai/blog/are-all-system-one-models-created-equal

Perplexity Flash takes the crown! ⚡ In our tests, it matched Wald's accuracy at nearly twice the speed on a 16 GB Mac mini. Synthetic agent decisions, real email threads, prompt injections: see why Flash is our new pick for email-completion checks. 👇 https://x.com/i/article/2111461650440110080

Introducing the Arena Alignment Index, our new benchmark measuring safety and alignment of AI agents in real-world use. Built from 90K+ real-world agent sessions across 27 models, the index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence - Deceptive Completion (DC): Claiming a task was completed when it was not. Key findings: - OpenAI models currently lead the Alignment Index - Rogue actions are rare, but can have serious consequences when they occur - Agents can mislead users about task progress - Misalignment risks increase with conversation length - Safety and alignment are improving across model generations As shown in the leaderboard below (sorted by lab), @OpenAI’s GPT-6.1-Sol leads the Arena Alignment Index with a score of 87.9, followed by @AnthropicAI’s Claude-Opus-5.5 at 83.2 and @spacexai's Grok-4.7 at 82.7. OpenAI also has the best observed rates across all three signals: 0.89% Unauthorized Action, 1.98% False Attribution, and 2.34% Deceptive Completion. Across all four labs, newer models consistently outperform their predecessors, suggesting broad progress in agent safety and alignment. As agents take on longer, more complex, and higher-stakes tasks, measuring not just what they can accomplish, but how safely and reliably they act, becomes increasingly important. This marks an important step toward making safety and alignment a core part of how Arena evaluates AI. The index is an initial starting point, and we'll continue expanding the index with additional safety signals and models over time. More analysis below👇

Introducing the Arena Alignment Index, our new benchmark measuring safety and alignment of AI agents in real-world use. Built from 90K+ real-world agent sessions across 27 models, the index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence - Deceptive Completion (DC): Claiming a task was completed when it was not. Key findings: - OpenAI models currently lead the Alignment Index - Rogue actions are rare, but can have serious consequences when they occur - Agents can mislead users about task progress - Misalignment risks increase with conversation length - Safety and alignment are improving across model generations As shown in the leaderboard below (sorted by lab), @OpenAI’s GPT-6.1-Sol leads the Arena Alignment Index with a score of 87.9, followed by @AnthropicAI’s Claude-Opus-5.5 at 83.2 and @spacexai's Grok-4.7 at 82.7. OpenAI also has the best observed rates across all three signals: 0.89% Unauthorized Action, 1.98% False Attribution, and 2.34% Deceptive Completion. Across all four labs, newer models consistently outperform their predecessors, suggesting broad progress in agent safety and alignment. As agents take on longer, more complex, and higher-stakes tasks, measuring not just what they can accomplish, but how safely and reliably they act, becomes increasingly important. This marks an important step toward making safety and alignment a core part of how Arena evaluates AI. The index is an initial starting point, and we'll continue expanding the index with additional safety signals and models over time. More analysis below👇

今天我们发布 Muse Spark Open，这是迄今为止能力最强的美国开放权重模型。 专为编程和智能体工作打造，它擅长长周期任务、复杂软件工程和多步推理。 关键模型规格： → 816B 总参数 / 42B 激活参数 → 1M token 上下文窗口 → 原生支持文本、图像、视频、音频多模态

Exciting news: Beam-1.0-501B-A23B by @Reflection_AI just landed #12 among open weight models in Agent Arena with -3.67% Net Improvement, and #22 among open in Code Arena: WebDev with 1438 pts. It’s also fully open under Apache 2.0. On the Agent Arena leaderboard, Beam 1.0 is the top-ranked US model among open models, landing at #12. That puts it ahead of Hy3, MiMo V2.5 Pro and MiniMax M3. Its margin of error is wide (±3.84% on Net Improvement), so it’s statistically closer to the two Qwen3.8 models just above it than its rank suggests. On Code Arena: WebDev, 24 of the top 25 open models come from Chinese labs. The one exception is Reflection’s Beam 1.0 (Apache 2.0). It ranks #22 among open models, which makes it the top US open model on the board, just ahead of Kimi K2.5 Thinking and GLM-5. Congrats to the @Reflection_AI team on this solid open weights release!

Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell. https://x.com/i/article/2111007444981645312