Microsoft Drops Claude Code, /usage Lands, and "You Guys Still Use IDEs?"
Claude Code & Anthropic Updates
/usage lands in Claude Code — per-skill, per-agent, per-MCP, per-plugin token attribution. Boris Cherny (@bcherny): "In the next version of Claude Code: run /usage to see a breakdown of which Skills, Agents, MCPs, and Plugins are using your tokens. CLI today, coming to Desktop next." 4.6k likes, 215k views, 288 replies. Key clarifications from Boris in-thread: it aggregates across all sessions (filterable by day/week); it counts tokens and attributes downstream tokens to the originating skill/plugin/MCP — not just raw input counts. He confirmed MCP support was added today and suggested users can ask Claude to write the resulting "don't use this tool" rules straight into settings.json to enforce pruning. Robin Cheng landed the most upvoted dissent: "Claude code being so expensive is kind of the bigger problem… Codex performs better and is way cheaper so I never even need to wonder where my token usages went." The community is treating this as the missing observability layer — "context window is the new stack trace; when an agent fails, ask what consumed context, not what code broke," one commenter wrote.
"OpenClaw is still around!" Mitsuhiko — a one-liner reacting to a fresh OpenClaw sighting. Two days after the rumored death, the open-source agent harness keeps showing up; Steipete RT'd Altay adding Cloudflare Containers support to crabbox, the tool he uses while reviewing OpenClaw PRs.
Side chats can be moved anywhere. Small CC UX detail retweeted by bcherny: the floating side-chat panel in Claude Code now drags freely across the workspace. Boris is shipping rapidly across the surface area.
Agentic Coding & Agent Harnesses
Matt Pocock's /handoff deep dive — an alternative to /compact. @mattpocockuk: "It's an alternative to /compact that gives you WAY more flexibility with your context window. Think of an idea, handoff to another agent to implement. Grill, handoff to prototype, handoff BACK." Video walkthrough; 1k likes, 68k views, 37 replies. Best follow-up from Daniel Escalante: a richer /resume companion that writes timestamped handoff files with YAML frontmatter capturing git snapshot (HEAD, dirty paths, ahead/behind), then runs a drift check against new commits when resumed — "the 'tried and rejected' section turns out to be the highest-leverage one." Matt also clarified the distinction from /fork: handoff is for clean state in the next session; fork preserves the original context for true parallelism. Separately he's wiring /improve-codebase-architecture to ping him a refactor suggestion every morning at 9 AM. Agents-as-cron.
Matt sharpens /tdd: no mirror tests. Tweet: "Adding a tweak to the /tdd skill: 'Do not add tests which simply restate the implementation. These provide zero confidence.' Getting sick of shit tests just to provide evidence of RGR." 766 likes, 50k views, 42 replies. When pressed on whether models even benefit from Red-Green-Refactor, Matt: "They are so much better with RGR, it's crazy." The most-cited horror story in-thread, from Vance Lever: "Had our AI agent write 4,000 tests in November. 100% pass rate. Head of engineering called it the most thorough coverage we'd ever seen. Pushed to production Friday. Entire auth layer stopped working. Turns out all 4,000 tests were asserting that the functions had been called." Ahmed Sulajman piled on: "I asked Codex to change the copy for something the other day, and it created a test to verify that the string in fact contains the new text."
swyx's 16-hour, 103-commit "slop → production-ready" skill. @swyx: "working on a 'take this vibecoded slop app and make it a production-ready, e2e tested, maintainable, parallelizable agent repo' skill. This thing ran for ~16 hours yesterday and made 103 commits all told and i ended up with exactly the same app but instead of fragile mvp it now looks like a codebase i can actually build on for the long run." 351 likes, 53 replies. The pull-quote of the thread, from Guilherme Lippert: "'Same app' is the tell. The user-visible product stayed the same, but the repo turned from a one-off demo into something future agents can keep touching without starting a rescue mission every time." GotNerfed countered with the metered-billing cost reality check: "a 16-hour 103-commit run on a usage-based plan is a real bill now. The 'let it run overnight' pattern only works while the vendor is subsidizing the tokens." (See the Microsoft section below — this is the live debate.)
"You guys still use IDEs?" Theo, quote-tweeting a droidbuilds survey post. 1.7k likes, 387k views, 269 replies — the biggest engagement of the day on agentic-coding Twitter. The reply field is a snapshot of where the discourse is: "Neovim + Claude Code" / "Cursor 80% / Codex 20%" / "these aren't IDEs, they're code editors" / "NLDE is the new IDE" / "half the day it's just Claude Code in tmux" / "AG is an IDE and NOT an IDE and I'm using both." The strongest pushback came from a Unity game-dev: "Unity is the IDE. You can't cd into a scene graph." Theo: "This is fair!" Vortex framed the maximalist take: "NGMI. 99% of actual work gets done in the chat/voice interface on Codex. Top-level builders are conductors in an orchestra — not actually playing each instrument."
"Codex is so much better." Adi Singh (retweeted by Theo): "I saw @theo's video about ditching Claude Code and thought 'it couldn't be that bad, right?' It is that bad. Codex is so much better." 190 likes, 15k views. The defenders held the line — Yash Khivasara: "For one-shot stuff, yeah Codex wins. For anything multi-file CC still handles context way better. Annoying rules but I haven't found anything cleaner for actual projects." mohbii landed the cleanest summary: "Claude code got too expensive and too inconsistent at the same time. Codex at least gives you predictable output even if it's not as creative. Reliability beats brilliance for actual shipping." The thread also surfaced a long, devastating Windows-Codex bug list from KorduGG — WSL crashing the app, browser/computer use broken, no automatic PR follow-up loop, freezing on launch, GPT 5.5 Pro labelled in the UI but not actually selectable. "Day by day I understand why Theo went ahead and made his own IDE."
Security & The Subsidy Cliff
Theo pins the GitHub breach: "really, really bad." @theo — the pinned post for the day. 569 likes, 97k views, 34 RTs, 31 replies. Video deep-dive. Replies range from "yep I'm done with GitHub, storing locally, looking for another solution" to "the real story is how many companies quietly built their entire auth stack on GitHub trust assumptions." Sufiyan Junaidi flagged the right frame: "The hard part is this isn't really a 'scan your code' problem — credentials got pulled from dev machines, not shipped code. Different attack surface than what most scanners handle." vinaykumar: "Repo contents become recon data even after keys rotate. Secrets are only the first pass." Feross thanked Theo for explaining the issue in plain terms; Sarah Gooding: "There's a real public service in helping everyone understand the nature of these attacks."
Microsoft cancels its internal Claude Code licenses. LLMJunky's "buy. a. gpu." tweet quotes a viral Hedgie post that's hard to ignore if true: "Microsoft canceled its internal Claude Code licenses this week after token-based billing made the cost untenable, even for a company with effectively infinite cloud resources. Uber's CTO sent an internal memo warning the company burned through its entire 2026 AI budget in just four months. American AI software prices have jumped 20% to 37%, and GitHub (owned by Microsoft) is dropping flat-rate plans for usage-based billing across its products." Hedgie's frame: "The AI subsidy era is ending in real time. The same company that put $13B into OpenAI and built the Azure infrastructure powering most of Anthropic's compute just looked at the bill from a competitor's coding tool and decided it was not worth paying." LLMJunky's personal stat: ~$3,000 in API spend this month alone — "that's 1/3 of a GPU right there." The most-cited counterpoint, from Dmitry Turmyshev: "Chinese models are getting very good, and with a 10x+ price difference, they start to make sense for a lot of tasks." Marvin Freeman: "Microsoft won't be the last, many will follow." This thread sits behind almost every other story today — the /usage instrumentation, the 16-hour-103-commit refactors, the IDE debate, the Codex-vs-CC fight, all sit on top of the same question: who pays for the tokens when the rebates run out.
Builders & Tools
Datasette Agent — Simon Willison's first alpha. @simonw: "The first alpha of Datasette Agent — a conversational AI assistant for Datasette that can answer questions about data in SQLite databases, and can be extended with plugins to add extra tools and features." Blog post: simonwillison.net/2026/May/21/datasette-agent/, interactive demo at agent.datasette.io. 131 likes, 11k views, 28 replies. When asked how this differs from "Claude Code + SQLite," Simon: "Run as a web app in your browser with the option to plug into many different models. Claude Code + SQLite is hard to beat one-on-one!" The thread immediately surfaced the right anxiety — Shivam Patel: "the second it reads a row some rando dumped data into, isn't that just prompt injection waiting to happen? Does the alpha default to read-only and gate writes?"
Composer 2.5 → #3 on the Artificial Analysis Coding Agent Index. Artificial Analysis, retweeted by leerob: "Cursor's new Composer 2.5 takes third on the Artificial Analysis Coding Agent Index and is ~10-60x lower cost than the higher-effort Opus 4.7 and GPT-5.5 variants above it. This release puts Composer among the leading coding agent models." Elon Musk replied "Amazing price for performance!" on Lee's feedback solicitation thread. CursorBench at cursor.com/evals.
MiniMax-M3 confirmed open-weight. Ryan Lee (RT'd by LLMJunky): "Yes, M3 will be open-weight. Coming soon." LLMJunky: "awesome. this guy cooks."
Pi is now native in the Entire CLI. Entire HQ, RT'd by Mitsuhiko: "Previously available as an external agent plugin, Pi now ships as a native integration. Your Pi sessions now connect directly to Entire checkpoints, commits, and session history." For Mitsuhiko himself, the better-known development today was getting Pi to call a Win32 API in 3 KB instead of 27 MB by ditching koffi for a hand-rolled MSVC shared lib — "never going to change, so can vendor 3kb of shared lib in the binary."
Datasette/Latent.Space — the AI infra unicorn issue. swyx's tone-of-the-day take: "everyone in ai infrastructure is finally getting filthy rich and it is so nice to see them succeed. not the sexy ai research stuff, just 'boring' infra." He linked the new AI News issue: latent.space/p/ainews-new-ai-infra-unicorns-exa. Best reply, from Mercurio Nevio: "The real moat wasn't models. It was surviving 4 years of being called 'just a wrapper' while quietly owning every GPU bottleneck." Also from swyx: a Daytona deep dive on agent-native compute — 60ms sandboxes, 50K startups in 75 sec, 850K daily runs, CLI > MCP, the end of localhost.
Industry & Discourse
LlamaIndex flattens all roles into "Member of Technical Staff." Jerry Liu: "As of yesterday, we made everyone within @llama_index research/engineering/product a Member of Technical Staff." His reasoning: "Eng, research, and product are collapsing into one role. Now that coding/project management are commoditized, every engineer is expected to own e2e outcomes and also 'know more things' across the stack." Flat structure, extreme autonomy, no senior-vs-junior gating. The pattern is the frontier-lab playbook (OpenAI/Anthropic/DeepMind) bleeding outward into application-layer companies — the same week LLMJunky is documenting Microsoft pulling Claude Code licenses, this is how lean upstarts plan to keep moving.
Karpathy → Anthropic; trq212 → Claude Code. Both confirmed earlier in the week; the engagement on Karpathy's post is still rolling. trq212's quote-RT of the Karpathy news: "the future is bright, lets get to work." Continuing the Anthropic ↔ SpaceX Colossus 2 narrative from yesterday's roundup.
Simon Willison on Google I/O 2026: "I don't have much to say about this year's Google I/O because I prefer to write about products that have shipped, not just 'coming soon' announcements — but here are some notes on Gemini Spark and Antigravity." Blog.
Mitsuhiko gets a public coaching moment. Badlogicgames captured it: "how it happened. @mitsuhiko: 'sensai, the code, it sparks no joy anymore.' pidalf: 'don't be a such a god damn fool.' @mitsuhiko: 'thank you. i can see the light again.'" The kind of week it's been on developer-AI Twitter.
Beyond Code
FTC: "active listening" microphone ad-targeting was a scam. Simon Willison: "I was VERY suspicious of this 'active listening' story when it first started circulating. Turns out it was a scam, they weren't targeting ads by listening through microphones at all." The FTC is making Cox Media Group and two other firms pay nearly $1M to settle deception charges over an AI-powered marketing service. Simon's original September-2024 take ages well.
OpenAI claims a breakthrough on Erdős's planar unit distance problem. RT'd skeptically by Jerry Liu: "proof too complicated, Claude help ELI5." OpenAI's framing: "For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids. An OpenAI model has now disproved that belief, discovering an entirely new family of constructions."