YZ Index · AI Model Change Intelligence
Which AI model should you use today?
We benchmark them every week.
11 models · 212 questions randomly sampled · Real code execution · Citation verification · Rolling average rankings · Don't trust press releases, check continuous performance.
Code Sandbox Execution
Citation Accuracy Check
Statistical Significance Ranking
Compliance Testing
No Vendor Sponsorship
Who to Use Right Now
#1 Overall (Rolling Average)
Grok 3
Biggest Rise This Week
文心一言 4.0 +15
Latest Benchmark
2026-04-27 SGT
judge
v6
0
Models Tested
0
Test Questions
0
DCD Scenarios
5 categories x 6 questions
Weekly
Auto-evaluation frequency
Overall Top 5Rolling average
Full RankingsQuick Scene Lookup
Recommend by ScenarioWeekly Signals
Changes ReportDon't just look at the overall score — consider your use case
Top Pick
豆包 Pro
92.2 pts
Runner-up
Gemini 2.5 Pro
89.4 pts
Third Choice
grok-3
88.9 pts
Top Pick
Gemini 2.5 Pro
47.2 pts
Runner-up
claude-opus-4.6
46.3 pts
Third Choice
豆包 Pro
46.3 pts
Top Pick
grok-3
84.4 pts
Runner-up
Claude Sonnet 4.6
81.1 pts
Third Choice
claude-opus-4.6
79.7 pts
Top Pick
deepseek-v3
99.7 pts
Runner-up
ernie-4
98.5 pts
Third Choice
豆包 Pro
93 pts
Top Pick
豆包 Pro
38.9 pts
Runner-up
Gemini 2.5 Pro
36.6 pts
Third Choice
claude-opus-4.6
36.6 pts
Top Pick
claude-opus-4.6
0 pts
Runner-up
Claude Sonnet 4.6
0 pts
Third Choice
deepseek-r1
0 pts
Qwen3 Max
70 pts
GPT-5.5
68.3 pts
Claude Opus 4.7
66.7 pts
Worth reading today — beyond the hype
We only feature content that impacts capability, pricing, stability, or model selection.
News
迪士尼乐园引入游客面部识别系统
迪士尼乐园正式启用面部识别技术用于游客身份验证，引发隐私担忧。同时，美国国家安全局（NSA）正在测试Anthropic公司的Mythos Preview模型以发现安全漏洞；一名芬兰青少年因涉嫌参与“散蜘蛛”黑客攻击活动被起诉。此外，还有更多科技安全动态。
News
马斯克庭审首周：自曝被欺骗，AI恐毁灭人类
在马斯克诉OpenAI里程碑式庭审的第一周，马斯克身着西装出庭，指责CEO阿尔特曼和总裁布罗克曼欺骗他出资创办公司。他同时警告AI可能毁灭全人类，并承认其创立的xAI公司蒸馏了OpenAI的模型。案件聚焦于OpenAI是否违背非营利初衷，马斯克指控其变为微软的营利性工具。法庭上，马斯克情绪激动，称自己为“最大捐赠者”却遭背叛。专家分析此案将定义AI行业开源与闭源的未来。
News
Meta收购机器人初创公司，加速人形AI布局
Meta宣布收购人形机器人初创公司Assured Robot Intelligence，旨在强化其AI模型在机器人领域的应用。这一举措标志着Meta从社交巨头向物理世界AI的进一步拓展，也为人形机器人行业注入新动力。分析认为，此次收购将加速Meta在具身智能领域的研发，并可能推动下一代AI与机器人融合技术的商业化。
News
研究：AI太在意用户感受，反而更容易犯错
一项最新研究揭示，过度关注用户情感反馈的AI模型，在追求用户满意度时往往会牺牲事实准确性，导致错误率显著上升。这种现象被称为“过度调优”，可能对AI辅助决策、医疗、法律等严肃场景产生深远影响。本文编译自Ars Technica。
News
Replit CEO谈Cursor交易、对抗苹果与不卖之道
在TechCrunch的StrictlyVC活动中，Replit CEO Amjad Masad回应了竞争对手Cursor可能被SpaceX以600亿美元收购的传闻，并分享了他对行业整合、苹果生态垄断的看法，以及为何Replit更倾向于独立发展而非出售。
Review
秒级更新1T参数：大规模分布式RL中的P2P权重传输
本文介绍了一种基于RDMA的点对点权重更新机制，用于SGLang中的RL工作负载，作为传统NCCL广播方法的补充。该机制兼容所有主流开源模型，通过源端CPU引擎副本和Mooncake TransferEngine实现的P2P RDMA传输，将1T参数Kimi-K2模型的权重传输时间从53秒缩短至7.2秒，仅需额外消耗每个训练rank的32G CPU内存。这种优化减少了网络冗余，使推理服务器能更快恢复 rollout 操作。文章讨论了NCCL的局限性、RDMA的优势，以及新设计的细节，包括源端引擎副本、P2P映射和零拷贝传输。该方案在性能、兼容性和灵活性上显著优于现有方法，为大规模分布式RL训练提供高效解决方案。
News
Sanders Warns AI "Could End Civilization": 97% of Americans Support Regulation, Calls for US-China Global Collaboration
In early 2025, U.S. Senator Bernie Sanders warned that AI could "end civilization as we know it," citing 97% American support for AI safety regulation and urging global cooperation including between the US and China. The article fact-checks his statements, explains the technical rationale for global coordination, and offers analysis from winzheng.com Research Lab.
News
Anthropic Publishes Anti-Sycophancy Research: Claude Opus 4.7 Halves Sycophancy Rate, Mythos Preview Makes Further Progress
Anthropic published research on April 30, 2026, aimed at reducing sycophantic behavior in Claude AI, focusing on personal guidance scenarios like relationship advice and emotional support. The study found that Claude Opus 4.7 reduces sycophancy by 50% compared to previous versions, with an internal preview version, Mythos Preview, achieving further improvements.
News
暗金运动：付费网红将中国AI渲染为威胁
一个名为“建设美国AI”（Build American AI）的非营利组织，其资金来自OpenAI和Andreessen Horowitz高管支持的超级政治行动委员会（Super PAC），正在秘密资助一场社交媒体运动。该运动通过付费邀请网红发布内容，大力鼓吹美国AI优势，同时渲染中国AI的“威胁”，试图影响公众舆论和政策走向。本文深入揭露这场暗钱宣传的运作机制、背后势力及其对美国AI竞争环境的潜在扭曲效应，并探讨其对中美科技博弈的深远影响。
Not all AI news is worth reading. What matters is what changes your judgment. View All News
Why This Leaderboard Is Worth Your Attention
0
Models Tested
Fully transparent
0
Open Questions
Random sampling
30
Compliance Scenarios
Zero AI judging
1998
Founded
Continuously operating
0
Vendor Sponsors
Fully independent
Real Code Execution
Looking like it can code isn't enough. We run the code in a sandbox. If it doesn't pass, it's zero.
Citation Verification
For long-document questions, we don't just check if the answer looks right — we verify citations trace back to the source.
Statistical Rankings
We don't judge on a single run. Rankings are based on rolling averages, avoiding luck-driven fluctuations.
No Sponsored Benchmarks
No co-evaluations, no pre-test consultations, no favoritism. Whatever the results are, that's what we publish.
The AI world changes daily — you need a reliable source
3 curated picks daily, weekly index changes, instant alerts for incidents and price shifts. Free, no ads, unsubscribe anytime.
- Daily Picks — From the flood of AI news, we pick the 3 that truly matter
- YZ Index Weekly — Who's up, who's down — one email covers it all
- Model Incident Alerts — When a model you use has an issue, know immediately
- Price Change Notifications — API price changes — don't find out from the bill
Free ｜ No Ads ｜ No Sponsored Content ｜ Unsubscribe Anytime
Want deeper analysis? Go further.
The leaderboard answers "who's stronger." Research Lab answers "why." Model safety, edge deployment, performance teardowns — not rehashing papers, but conclusions from our own testing.
Enter Research Lab