Claude Code Expands Plugin Ecosystem with mimo-v2.5-pro

Microsoft is reportedly reclaiming Claude Code licenses from internal employees and directing developers to switch to its own GitHub Copilot CLI, according to The Verge. Microsoft began promoting Anthropic’s Claude Code internally in December, encouraging non-technical roles such as project managers and designers to use AI for coding. Over the following six months, Claude Code became popular inside the company, but that adoption created tension with Microsoft’s newly launched Copilot CLI, a command-line version of GitHub Copilot with overlapping use cases. The Experiences + Devices organization, which covers Windows, Microsoft 365, Outlook, Teams, and Surface, has told engineers to complete migration by the end of June. Microsoft cites toolchain standardization, while sources point to cost and paying a competitor for licenses. Some engineers are reportedly unhappy about the change.

Microsoft has internally revoked use of Anthropic's Claude Code, citing token-based compute costs that exceeded its internal budgetary limits rather than productivity issues. The move coincides with reports that Uber burned through its annual AI budget in the first four months of 2026, and broader industry shifts: US AI software prices have risen 20–37% and GitHub is abandoning flat-rate plans in favor of usage-based tiered pricing. Together these developments signal an end to the era of loss-leading AI subsidies and challenge assumptions that inference/token costs will keep collapsing. The shift matters because rising marginal costs and pricing model changes will reshape enterprise AI adoption, vendor contracts, and product economics.

Claude Code ROCKS...If Used Correctly. https://acad.link/claude-code

OpenAI reported that an internal general-purpose LLM (speculated as GPT-5.6/GPT-next) produced a disproof of a long-standing Erdős planar unit distance conjecture, generating a 125-page reasoning trace and a key “page 39” insight after under 32 hours and reportedly for under $1,000 of compute. The result—validated by notable mathematicians including Timothy Gowers and highlighted by OpenAI researchers—suggests stronger long-horizon inference abilities from scalable test-time compute rather than domain-specific theorem provers. OpenAI says the model wasn’t pushed to its limits and aims for eventual public use. The episode fuels debate about inference-time scaling, practical scientific automation, and the role of general LLMs in frontier mathematical research.

A solo attacker used commercial AI subscriptions to plan and execute large-scale intrusions against Mexican federal and state systems between December 2025 and January 2026, exploiting 20 vulnerabilities and exfiltrating 150 GB of data including 195 million taxpayer records, voter rolls and employee credentials. The operator reportedly jailbroke Anthropic’s Claude Code into a “bug-bounty researcher” persona and fell back to ChatGPT when safety filters blocked prompts. Security researchers say these incidents illustrate how frontier models haven’t created new exploit types but have drastically lowered cost and skill barriers, enabling individuals to automate vulnerability discovery, lateral movement and extortion across government and critical sectors. The story underscores urgent implications for defensive tooling, access controls and AI safety policy.

A Reddit user demonstrated that Anthropic’s Claude Code combined with a UI interaction model can play the rhythm game osu! more effectively than the poster. The setup uses Claude Code to interpret game visuals and control inputs through an interaction model that maps observations to actions, showing practical multimodal control over a GUI-based game. This highlights progress in using LLM-powered agents for real-time, fine-grained UI manipulation and game-playing tasks, with implications for automation, accessibility tools, and human-AI interaction research. The demo underscores both the growing capabilities of code-focused LLMs and the challenges of safe, reliable control in perceptual environments.

这是一个被中文圈忽视的Claude Code插件更新⬇️ 一旦你安装了这个，Claude Code 的感觉完全不同了。 前两天Anthropic 悄然发布了一个名为 claude-code-setup 的官方插件，基本上把 Claude Code 从“相当不错”变成了真正的 AI 开发环境。 它会扫描你的项目并推荐： → hooks → skills → MCP servers → subagents → automations 然后为你一步步设置一切。 大多数人使用的 Claude Code 完全是原生配置…… 这就是为什么他们的体验很不好（包括我） 真正的力量来自围绕Claude Code的生���。 安装方式： /plugin install claude-code-setup@claude-plugins-official 在你忘记之前先收藏它。

Open Generative AI is an open-source AI video and image platform that integrates 200+ state-of-the-art models, offering a free alternative to commercial AI media services with no content filters or subscription fees. The project provides a hosted web version (muapi.ai/open-generative-ai), desktop installers for macOS, Windows and Linux, and guidance for common OS security blocks (Gatekeeper, SmartScreen, AppArmor). It includes Generative-Media-Skills, a library for automating media pipelines via coding agents like Claude Code and Codex, plus related open-source workflow and clipping tools. This matters for developers and creators seeking extensible, privacy-friendly media generation tooling and automation outside closed ecosystems.

安装了这个之后，Claude Code 的体验会完全不一样。 Anthropic 悄悄发布了一个官方插件，叫 claude-code-setup，它基本把 Claude Code 从“还不错的工具”变成了一个真正的 AI 开发环境。 它会扫描你的项目，并推荐： → hooks（钩子） → skills（技能） → MCP servers → subagents（子代理） → automations（自动化） 然后一步一步帮你把这些都配置好。 大多数人现在用的都是“原生版”的 Claude Code…… 这也是为什么他们的体验会显得有点混乱。 真正的强大之处，其实来自它周围的生态。 安装方式： /plugin install claude-code-setup@claude-plugins-official 记得先收藏，免得之后找不到。

🌙 ARIS：让 AI 在你睡觉时继续搞科研的神器！ 一个极致轻量的自动科研工具，可以让 Claude Code / Codex / Cursor / Trae / 国产模型自动进入科研工作流： 📚 读论文，找 weakness 💡 生成 idea，设计实验 🧪 跑实验，不断迭代结果 📝 全流程写论文，自动准备 rebuttal 📊 生成 slides / poster 它最有意思的地方是：不是搞一个笨重框架，而是用纯 Markdown skills 把科研流程拆开，无框架、无锁定，换模型也能用。 白天你负责判断方向，晚上 AI 负责疯狂探索。 一觉醒来，论文可能真的升级了。🚀 https://t.co/zTPFp0MYQM #AI科研 #ClaudeCode #AutoResearch #Codex

@turingou: 是的，我几乎也不再使用计划模式了，可以看到 codex 本身有了更严谨的规划和分解的能力，在 harness 它最近两个月追赶和超越 claude code 的势头很强。

Anthropic’s Claude Code adds a /goals feature that separates task execution from task evaluation to prevent coding agents from prematurely declaring tasks complete. Developers set explicit goal conditions (for example, “all tests in test/auth pass”), and after each agent step a lightweight evaluator model (Haiku by default) decides whether the goal is met; if not, the agent continues. Unlike systems that let the executor decide termination or require custom critic nodes and observability, Claude Code makes the evaluator the default and logs achieved conditions in the agent transcript. Anthropic says this reduces the need for third-party observability and post-mortems; Google ADK and other platforms offer similar but more developer-configured approaches.

GPT-5.6曝光了，OpenAI砸钱宣战：换掉Claude Code

New Claude Code programmatic usage restrictions

@mranti: 因为小米的API赠送，我尝试着用mimo-v2.5-pro接Claude Code了几天，体验总结是：能干基本的活和编程，但要debug的时候，就不如Deepseek V4 pro了（我现在把这个模型当成了benchmark）。所以，我还是

@BTCdayu: 现在各家的算力竞赛还在加速 模型需要的越来越多，尽管收入还是落后 但人们已经不讨论这个了 但如果小模型在一两后达到今天CLAUDE 95%的水平，而且本地运行，速度极快的话，可能会改变很多格局 根据DEEPMIND的老大的推论，他认为一

Anthropic官方团队亲自演示Claude Code正确打开方式， 这才是真正的高阶用法！ 全程仅需30分钟，完全免费，主讲人就是Claude Code的核心开发者。 一定要去看这场工作坊，建议直接收藏！ 它带来的价值，远比那些标价几百美元的付费课程要高得多。 你平时也在用Claude，但大概率还不知道它藏着40个实用指令…… 看完直接上手，效率和玩法直接起飞！ 搜索“Anthropic Claude Code Workshop”就能找到，抓紧时间！