Latest

Every published item, in order · 140 items

Blogs News Posts Papers Papers →

Nov 8

Papers Builders06:08

Chinese contentStanford 论文：更多信息会让部分 LLM 谈判智能体收益下降

Stanford 测了 10 个 LLM 谈判，发现给模型更多信息反而可能让 Claude Opus 5.5 收益从 94% 掉到 64%，选模型时得连信息量一起测。

#LLM #智能体 #Stanford #Claude

rohanpaul_ai@rohanpaul_aiSource

Nov 7

Models US vs China23:47

Chinese content网络攻防基准：开源模型 DeepSeek V4.1 Flash 追近 GPT-6

Vals AI 测了 GPT-6 和 DeepSeek V4.1 的网络攻击能力，开源模型花十分之一的钱就追到只差 8 个百分点，做安全的得看看。

#GPT-6 #DeepSeek #Vals AI #AI安全

官方账号Decoder@Matthias BastianSource

Models US vs China03:26

Chinese contentAA-PersonalityBench：Qwen3.8 27B 回复最长，约为 GPT-6 Sol 的 6 倍

同一道题，Qwen3.8 27B 能写一千词，GPT-6 Sol 只回一百多，这份字数榜把各家模型的啰嗦程度排得明明白白。

#Qwen3.8 #GPT-6 #K2 Horizon #Artificial Analysis

官方账号Artificial Analysis@ArtificialAnlysSource

Industry US vs China71°02:50

Chinese contentRamp 数据：企业 token 用量涨 65%，支出仅涨 1.5%，Anthropic IPO 估值预期降至 1.2 万亿至 1.3 万亿美元

企业 token 用量暴涨但支出几乎没涨，这份数据直接动摇 AI 高增长叙事，还牵出 Anthropic IPO 估值缩水，做 AI 生意的都该看看。

#Anthropic #OpenAI #Ramp #IPO

宝玉@doteySource

Models BuildersSelected00:56

Nace AI's 27B Drex Max Picks Instead of Generating Tokens

Routing thousands of micro-decisions to a pick-only model could cut agent orchestration costs by orders of magnitude where the answer space is a finite option list, not free text. Builders running multi-step agent pipelines can plug it in as a cheap decision layer while keeping a general model for generation. The gap on HLE shows it is a narrow tool, not a replacement for reasoning models.

#Drex Max #Nace AI #GPT-6 #Claude Opus 5.5

rohanpaul_ai@rohanpaul_aiSource

Products Builders00:53

Chinese contentDaybreak 面向安全团队开放：GPT-6 Sol/Luna 找漏洞，GPT-5.6 Cyber 测试攻击

OpenAI 给搞安全的朋友出活儿了：GPT-6 找 bug，GPT-5.6 Cyber 模拟攻击，红队还能申请 GPT-6 Cyber，得报名。

#Daybreak #GPT-6 #GPT-5.6 Cyber #AI安全

Vaibhav Srivastav@reach_vbSource

Models Builders00:45

Chinese contentNace AI 开源多模态决策模型家族 Drex 并发布技术报告

Nace AI 开源了决策模型 Drex，权重代码都放出来了，Drex Max 在几个决策基准上能跟 GPT-6 Sol 掰手腕，想本地部署可以试试。

#Drex #Nace AI #开源模型 #多模态

Aran Komatsuzaki (论文推介)@arankomatsuzakiSource

Nov 6

Models BuildersSelected22:03

Nace AI releases Drex open decision models up to 27B

The interesting part is the interface shift: instead of generating text, these models emit calibrated probabilities over fixed options, which suits scoring, routing and agentic decision points where latency matters. Builders assembling agent pipelines or classification backends can try the weights on Hugging Face or read the paper at arxiv.org/abs/2611.07854. Compared with calling Claude Opus 5.5 or GPT-6 Sol, Drex trades text generation for sub-200ms structured decisions under an open licence.

#Drex #Nace AI #Claude Opus 5.5 #open-source models

elvis@omarsar0Source

Nov 4

Papers US vs China01:32

Chinese contentAllen AI 发布 StateBench：GPT-6 Sol 与 Claude Opus 5.5 状态追踪准确率仅 13%

Allen AI 做了个新基准 StateBench，专门考模型能不能跟上长文本里的状态变化，结果 GPT-6 Sol 和 Claude Opus 5.5 在最难题目上准确率只有 13%，挺扎心的。

#StateBench #GPT-6 Sol #Claude Opus 5.5 #Allen AI

官方账号Allen AI (Ai2)@allen_aiSource

Models US vs ChinaSelected00:45

Kimi K3.1 Pro shows largest personality shift in AA-PersonalityBench

A personality rewrite this large between consecutive versions means teams tuning prompts or evals around Kimi K3's behavioural profile may see different outputs after upgrading. Developers building character-driven agents or customer-facing assistants should re-run behavioural evaluations before switching to K3.1 Pro. The benchmark also offers a reusable method for tracking such shifts across any model lineage.

#Kimi #Moonshot #AA-PersonalityBench #benchmark

官方账号Artificial Analysis@ArtificialAnlysSource

Oct 30

Papers13:37

Chinese contentDecision 与 System One 模型在风湿病学考试题和诊断病例上的评测

拿考试题和真实诊断病例测了一批医疗 AI 模型，结果反差挺大：考试能上 93%，真病例却掉到 50%，高置信答案也没那么靠谱。

#System One #GPT-6 #RheumaMIR #RheumBench

arXiv: OpenAI@Alfredo Madrid-García et al. (6)Source

Oct 29

Models US vs China04:18

Chinese contentAA-PersonalityBench 测评：GPT-6 与 Grok 回答最直接

想挑个说话不绕弯子的模型？这份榜单直接量化了各家的“直球”程度，GPT-6 Sol 和 Grok 系列得分最高。

#GPT-6 #Grok #AA-PersonalityBench #推理模型

官方账号Artificial Analysis@ArtificialAnlysSource

Products Builders72°00:16

Chinese contentGitHub 发布 Copilot Fusion，动态调度 Anthropic 与 OpenAI 模型

GitHub 新出的 Copilot Fusion 会自动在 Claude 和 GPT 模型里挑着用，跑分只比 Opus 5.5 低 2 分，成本却便宜一半。

#GitHub Copilot #Copilot Fusion #Claude #OpenAI

官方账号Artificial Analysis@ArtificialAnlysSource

Oct 28

Models US vs China10:33

Chinese content谷歌更新 Android Bench：Gemini 4 Argon 编程成绩落后 Claude 与 GPT

谷歌自家基准把自家 Gemini 4 Argon 排到第五，通过率才 21.3%，GPT-6.1 Sol 的性价比数据对安卓开发者选模型很实用。

#Gemini #Claude #GPT-6.1 Sol #Android Bench

IT之家Source

Models US vs China72°10:06

Chinese contentOpenAI 公示 GPT-6.1 Terra 安全测试结果：整体风险下降，破坏性行为标记上升

OpenAI 公布 GPT-6.1 Terra 的安全测试细节，整体风险明显下降，但 13.75% 的样本出现未授权操作，部署 Agent 前值得看看这些数据。

#GPT-6.1 Terra #OpenAI #AI安全 #智能体

IT之家Source

Oct 27

Papers Builders09:23

Chinese contentSecPRBench：用真实漏洞提交评测智能体代码审计能力

想知道 Claude 或 GPT 审代码到底能不能抓到真漏洞？这篇基准测了八款模型，265 个来自 Firefox、OpenSSL 的真实漏洞提交，结果差异挺大。

#SecPRBench #Grok 4.6 #GPT-6 #代码审计

arXiv: OpenAI@Yihan Yang et al. (8)Source

Oct 24

Models US vs China71°06:56

Chinese contentKimi K3.1 Pro 登顶 Artificial Analysis 网络防御开源模型榜首

Moonshot 的 Kimi K3.1 Pro 在网络防御基准拿了开源模型第一，CyberGym-E2E-AA 得分 95% 连 GPT-6 Sol 都被超过，不过单任务成本反而变贵了，这篇把各家分数和价格都列得很清楚。

#Kimi #Moonshot #DeepSeek #推理模型

官方账号Artificial Analysis@ArtificialAnlysSource

Models US vs China74°01:26

Chinese contentDeepSeek V4.1 Flash 1017 登顶开放模型网络防御榜单

DeepSeek 新版开源模型在网络防御榜单上拿了开放模型第一，总分只差 GPT-6 受限版 12 分，价格还便宜 4 倍。

#DeepSeek #GPT-6 #Grok #开源模型

官方账号Artificial Analysis@ArtificialAnlysSource

Oct 23

Models US vs ChinaSelected06:51

AA-PersonalityBench links model slop to intelligence gaps

Slop output is only loosely tied to raw capability, so model selection for writing or content pipelines should weigh the Slop ranking separately from the Intelligence Index. Teams building customer-facing copy or summarisation tools can use these two scores as a joint filter instead of assuming a smarter model reads better. Qwen3.8 Max ranking high on both lists is a useful data point when weighing it against Gemini 4 Argon or GPT-6 variants.

#GPT-6 #Qwen3.8 Max #Claude #Gemini

官方账号Artificial Analysis@ArtificialAnlysSource

Oct 20

Models US vs China74°02:47

Chinese contentDeepSeek-V4.1-Flash 登顶 Agent Arena 开源模型第一

DeepSeek 这版 Flash 模型在 Agent Arena 拿下开源第一，跑任务成本只有 GPT-6.1 Sol 的四分之一左右，写网页还追平了闭源大模型，做智能体的可以看看。

#DeepSeek #Agent Arena #智能体 #开源模型

lmarena.ai@lmarena_aiSource

Oct 19

Models US vs China81°03:34

Chinese contentDeepSeek-V4.1-Flash-1017 登顶 Code Arena 开源模型榜首

DeepSeek 新版编程模型把 Kimi K3 挤下开源第一，成本只有十分之一，价格没涨，做 Web 开发的可以直接去 API 试试。

#DeepSeek #Kimi #Code Arena #开源模型

lmarena.ai@lmarena_aiSource

Oct 15

Papers BuildersSelected12:52

BizBench Benchmark Tests Agent Execution Validity in Payments Workflows

The benchmark shows that outcome-only evaluation masks invalid execution, since models can hit the final goal while violating constraints mid-run — a gap that matters most in regulated domains like payments where intermediate errors carry financial or compliance consequences. Agent builders can use the 50-task suite and deterministic checks to measure constraint compliance and tool-call overhead, not just task success. Teams choosing models for workflow automation get a direct comparison of frontier and open models under identical conditions.

#BizBench #agents #benchmark #Claude Opus 5

arXiv: DeepSeek@Enhan Zhao et al. (9)Source

Oct 14

Models US vs China21:50

Chinese content通用模型封锁网络安全能力，Artemis 推出防御专用模型 Orion-1

通用模型锁死安全能力，Artemis 的 Orion-1 专做防御，还跑赢了 Claude 和 GPT。

#Orion-1 #Artemis Security #网络安全 #AI安全

官方账号Decoder@Jonathan KemperSource

Models US vs China71°08:35

Chinese contentllm-gemini 0.35 支持 Gemini 4 Argon 模型

Gemini 4 Argon 今天上线，llm 插件已经支持了。定价对比讲得很细，还搬了 Artificial Analysis 的成本数据，选模型前可以看看。

#Gemini 4 Argon #llm-gemini #GLM-5.3 #Claude Opus 5.5

官方账号Simon Willison’s WeblogSource

Oct 11

How-to Builders23:53

Chinese contentAI团队消耗大量token但质量提升有限

Vals AI发现AI团队协作成本高但收益小，Anthropic数据进一步验证了这一现象。

#AI团队 #GPT-6 #Claude #token效率

官方账号Decoder@Matthias BastianSource

Models US vs China01:15

Chinese content微软发布 Microsoft-Decision-1 决策模型，准确率超 GPT-6 Luna Decision

微软把 Qwen3.5-9B 改成小决策模型，分类任务准确率干到 83.5%，跑分类便宜 200 倍，做内容审核的可以试试。

#Microsoft-Decision-1 #Qwen3.5-9B #OpenAI #决策模型

Blognone@mkSource

Oct 10

Models Builders07:21

Chinese content微软发布 Microsoft-Decision-1 决策模型

微软新决策模型响应速度超快，比 GPT-6 Sol 快 35 倍，价格还便宜。

#Microsoft-Decision-1 #Qwen3.5-9B #决策模型 #OpenRouter

IT之家Source

Models Builders07:05

Chinese content微软发布 Microsoft-Decision-1：基于 Qwen3.5-9B 的单次前向传播决策模型

微软基于 Qwen3.5-9B 出了个决策小模型，专管 Agent 流程里的判断打分，比 GPT-6 Sol 快 35 倍，输出还免费。

#Microsoft-Decision-1 #Qwen3.5 #Agent #决策模型

AIGCLINK@aigclinkSource

Models US vs China06:56

Chinese contentMicrosoft-Decision-1 上线 OpenRouter，主打快速决策

微软把决策类小模型做成了专项产品，比 GPT-6 Sol 快 35 倍还输出免费，做分类和判断任务的可以试试。

#Microsoft-Decision-1 #OpenRouter #Qwen3.5 #推理模型

OpenRouter@OpenRouterAISource

Models US vs ChinaSelected73°06:26

Microsoft-Decision-1 AI model released

This specialized decision model addresses a critical inefficiency in AI applications - replacing expensive general-purpose models for simple choice-based tasks. For developers in Southeast Asia, it offers significant cost reductions (up to 200x) and speed improvements (up to 100x) for tasks like content moderation, intent recognition, and AI output evaluation. The model's availability through Microsoft Foundry and OpenRouter makes it immediately accessible for production use cases across the region.

#Microsoft-Decision-1 #Qwen3.5-9B #decision models #model routing

宝玉@doteySource

Models Builders03:57

Chinese contentMicrosoft 发布 Microsoft-Decision-1 快速决策模型，已在 Foundry 上线

微软用 Qwen3.5-9B 做了个专门快速打分决策的模型，比 GPT-6 Sol 快 35 倍，Foundry 上就能用。

#Microsoft-Decision-1 #Qwen3.5 #Microsoft Foundry #GPT-6 Sol

TestingCatalog@testingcatalogSource

Oct 9

Models Builders79°05:27

Chinese contentOpenAI推出GPT-6.1 Sol架构

OpenAI新发布的GPT-6.1 Sol不仅价格减半，处理长提示速度也更快了。

#GPT-6.1 Sol #OpenAI #长文本处理

官方账号Epoch AI@EpochAIResearchSource

Models US vs China71°03:38

Chinese contentArtificial Analysis 网络安全指数新增受信访问模型，GPT-6 Sol（Daybreak Blue）登顶

OpenAI 给受信访问的 GPT-6 Sol（Daybreak Blue）几乎没有网络安全防护，跑分直接登顶还比 Grok 4.7 便宜好几倍，值得看看细节。

#GPT-6 Sol #OpenAI #Grok 4.7 #Artificial Analysis

官方账号Artificial Analysis@ArtificialAnlysSource

Papers BuildersSelected00:46

Arena study finds GPT-6 Luna and Astra misattribute user statements

False attribution matters more than raw accuracy for agent deployments, because a model that invents user consent or intent can trigger actions the user never authorised. Teams building approval-gated agents, audit logging or compliance workflows should test for this failure mode separately from hallucination benchmarks, since the two do not track together across the GPT-6 family.

#Arena #GPT-6 Luna #GPT-6 Astra #AI safety

lmarena.ai@lmarena_aiSource