Hello, AI Explorer!
Good Saturday. This week: Anthropic secured 220,000 GPUs and committed $200 billion to Google Cloud in the same five days, GPT-5.5 Instant quietly replaced GPT-4o as ChatGPT's default model, Claude's agents now grade their own work between sessions.
Welcome back to your weekly AI digest: human-curated, zero fluff.
Every Wednesday and Saturday, we bring you:
- The week's most important AI developments (no fluff)
- Hand-picked AI tools actually worth your time
- Expert insights on what these changes mean for you
Let's dive into this week's discoveries! ⚡
🔥 This Week in AI
🔥 This Week in AI |
May 9, 2026 |
| 🔌 | Anthropic leased SpaceX's Colossus 1 -- 220K+ Nvidia GPUs, Claude Code limits now doubled |
| 💰 | Anthropic commits $200B to Google Cloud over five years -- 40% of Cloud's disclosed backlog |
| ⚡ | GPT-5.5 Instant is now ChatGPT's default model -- 52.5% fewer hallucinated claims than GPT-4o |
| 🧠 | Claude Managed Agents now "dream" -- agents study past sessions and grade their own work between runs |
| 🌐 | OpenAI Codex now runs inside Chrome -- parallel agent across every tab, no session hijacking |
| 📋 | Claude lands in Microsoft 365 -- Excel, Word, PowerPoint, and Outlook with one shared conversation context |
| 🏥 | OpenAI o1 outperforms ER doctors in Harvard study -- 67% accuracy vs 55% for emergency physicians |
| 🔍 | Google AI Search now quotes Reddit and forum posts with creator handles and direct source links |
| ⚡ | 4 AI Tools This Week: Fireflies, Descript, NotebookLM, Rows -- all inside. |
Lead Story · Anthropic
Anthropic Locked In 220,000 GPUs and a $200B Cloud Commitment in One Week
Two deals landed inside five days. First: Anthropic leased SpaceX's Colossus 1 data center in Memphis — over 220,000 Nvidia GPUs, making it one of the largest single AI compute deployments outside of the hyperscalers. The lease is short-term but the signal is not: Anthropic needed more raw compute than any existing cloud provider could hand over immediately.
Then came the Google Cloud deal — $200 billion over five years. That figure represents roughly 40% of Google Cloud's entire disclosed customer backlog. For context, that's more than most Fortune 100 companies spend on IT infrastructure across a decade. Anthropic isn't just renting servers; it's becoming one of Google Cloud's defining anchor tenants.
The side effect users noticed immediately: Claude Code API rate limits were doubled. More compute flowing to Anthropic translates directly to fewer "rate limit reached" errors mid-task. If you've been hitting walls while running Claude Code on longer projects, this week should feel noticeably different.
What this means for you
Anthropic now has the infrastructure to run at a scale that matches OpenAI's. Expect Claude model releases to accelerate — and Claude Code capacity to keep improving through the rest of 2026.
OpenAI
GPT-5.5 Instant Is ChatGPT's New Default 52% Fewer Hallucinations Than GPT-4o
OpenAI swapped the default ChatGPT model this week without much ceremony. GPT-5.5 Instant now loads when you open a new chat replacing GPT-4o as the baseline experience for every user across Free, Plus, and Pro tiers.
The headline number: 52.5% fewer hallucinations compared to GPT-4o on OpenAI's internal benchmarks. Response latency is also down. OpenAI positions this as the "everyday reasoning" model faster than o3, cheaper to run, and meaningfully more accurate than the 4o line it replaces.
o3 and o4-mini remain available for tasks that need extended reasoning. But for most daily use drafting, summarizing, quick Q&A GPT-5.5 Instant is now what ChatGPT is.
Read the full story →
Also This Week
Anthropic
Claude Managed Agents Now "Dream" Between Sessions
Anthropic's Managed Agents pipeline added a dreaming phase — agents review logs from completed sessions, identify where they went wrong, and generate self-improvement notes before the next run. No human in the loop. The agents grade their own work and store the corrections. Over repeated cycles, task accuracy improves without any retraining.
Read more →OpenAI
OpenAI Codex Now Runs Inside Chrome — One Agent Across Every Tab
The new OpenAI Codex Chrome extension puts a coding agent directly in the browser. It can read the active page, run code in a sandboxed environment, and work across multiple tabs in parallel. No switching to the API playground — it runs inline while you browse. Early users are using it to scrape, debug frontend code live, and automate repetitive browser tasks without leaving the page.
Read more →
⚡ AI Tools of the Week |
May 9, 2026 |
|
🌟 Hero Pick — Meeting AI Joins your Zoom, Meet, or Teams call automatically. Records, transcribes, and generates a full action-item summary within minutes of the call ending. The search across past meeting transcripts is genuinely useful — type a name or topic and pull up exactly what was said three meetings ago. With Claude now inside Microsoft 365, meeting AI is having a moment. Fireflies is the standalone version that works across every platform at once. Try Fireflies → |
|
💜 Staff Pick — Video & Podcast Edit video and audio by editing the transcript. Delete a sentence in the text, the clip disappears. Record an overdub of your own voice to fix a stumble without re-recording. Descript's AI Screen Recorder and Clip tool also turn long recordings into short shareable clips automatically. The fastest path from raw recording to published content — no timeline scrubbing required. Try Descript → |
|
🆓 Free Pick — Research & Study Upload PDFs, docs, YouTube links, or audio files. NotebookLM builds a private knowledge base from your sources and lets you ask questions, get summaries, and generate study guides — all grounded in your actual documents, not the open web. The Audio Overview feature turns any source into a two-host podcast-style conversation. Completely free. One of Google's best under-discussed releases of the past year. Try NotebookLM → |
|
💎 Hidden Gem — Spreadsheet AI A spreadsheet with an AI analyst built in. Ask it questions about your data in plain English, generate charts, and connect live data sources (Stripe, Google Analytics, HubSpot) without writing a single formula. Relevant this week given OpenAI quietly added ChatGPT to Google Sheets — Rows does the same thing but with deeper integrations and cleaner UX, and it's been doing it longer. Worth a look before the ChatGPT version locks in as your default. Try Rows → |
⚡ Quick Hits
🏥 OpenAI o1 outperforms ER doctors on Harvard diagnostic study — 67% accuracy vs. 55% for physicians on 300 complex emergency room cases. The study controlled for physician specialty and time pressure. Read →
🔍 Google AI Search now cites Reddit and forum posts by creator handle — AI Overviews started attributing user-generated content with the original poster's username. First time UGC has been surfaced this directly inside Google's AI results. Read →
💻 Perplexity launches desktop app for Mac — Full computer use capability: browse the web, write files, and control apps without a browser tab. Positioned as a direct competitor to Operator and Claude Computer Use. Free tier included. Read →
📋 Claude is now available inside Microsoft 365 — Excel, Word, PowerPoint, and Outlook. Announced as part of Microsoft's Copilot+ expansion. Adds a second AI option inside the apps where most knowledge workers already spend their day. Read →
🎮 Google DeepMind's AI agent beats veteran EVE Online players — The agent, trained entirely through self-play, won multi-hour strategic sessions in one of the most complex persistent-world games. DeepMind says it's a long-horizon planning benchmark, not a gaming demo. Read →
📖 Worth Reading This Weekend
Three longer reads if you want to go deeper
The Mythos AI data deal and the White House connection
~6 min read
A detailed look at how a little-known AI data company landed government contracts and the political relationships behind the deal. Important context for anyone following the AI policy and procurement space.
Read →OpenAI's $10B private equity joint venture — what they're actually building
~8 min read
Details on the structure of OpenAI's new PE partnership and what the capital is earmarked for. Spoiler: it's less about model training and more about vertical integration into specific industries.
Read →Sub-Q raises $12M for 12M-context window memory for enterprise agents
~5 min read
One of the more interesting infrastructure bets this week. Sub-Q is building persistent long-context memory as a standalone service — so any agent can "remember" across sessions without embedding the full history in the prompt. Relevant if you build with LLMs.
Read →That's all for this edition! What tools are you most excited to try?
Reply to this email - we read every response and use your feedback to improve future editions.
See you on Wednesday,
The AI Tool Discovery Team
Feedback
What'd you think of today's newsletter?
Vote below to let us know how we're doing.