SpaceX Buys Cursor for $60B,, Government Shuts Down Fable 5, 📉 ChatGPT Below 50%
DeepSeek's $7.4B raise, Cursor's agent-native GitHub rival, agentic code churn up 861%, and 5 new model releases.
This Week’s Big Stories
Must-Read
US Government Forces Anthropic to Shut Down Fable 5 and Mythos 5 — The Commerce Department issued an export-control directive forcing Anthropic to suspend access to its most capable models after Amazon CEO Andy Jassy’s conversations with US officials flagged jailbreak vulnerabilities. Amazon researchers had used targeted prompts to get Fable 5 to provide actionable cyberattack information. What followed was the AI industry’s most consequential government intervention to date: a week-long arc spanning employee protests, 150+ cybersecurity experts signing an open letter calling for the restrictions to be lifted, and active negotiations between Anthropic and Trump officials to restore access. Anthropic also paused its controversial token-based SDK billing changes amid the crisis. The deeper question — whether Anthropic’s safety capabilities are a superpower or a liability — remains open.
SpaceX Acquires Cursor for $60 Billion — SpaceX is buying the AI code editor in an all-stock deal, the largest acquisition of a developer tool in history. Cursor crossed $1B ARR in November 2025 and has become the default environment for AI-assisted development at hundreds of companies. The deal is expected to close in Q3, subject to regulatory approval. What SpaceX plans to do with an AI code editor — whether this is about internal tooling for Starship software, a platform play, or a talent acquisition at scale — remains unclear.
ChatGPT Market Share Drops Below 50% for First Time — ChatGPT is still the most popular AI chatbot, but users are increasingly willing to switch. Gemini, Claude, and Grok are all gaining share. The timing is notable: the same week, Noam Shazeer — co-author of the “Attention Is All You Need” paper that started the transformer era — announced he’s leaving Google for OpenAI. That’s the kind of hire you make when you can feel the ground shifting beneath you.
Cursor Launches Origin — an Agent-Native GitHub Competitor — Cursor announced Origin, a Git-compatible code forge built for AI agents cloning, branching, committing, and rebasing in parallel. Where GitHub was built for human-scale collaboration, Origin is designed for agent-scale — hundreds of agents working a codebase simultaneously. Same week, Cursor announced a 1.5T-parameter model trained from scratch on 100K+ GPUs, with capabilities extending beyond coding. Two signals pointing the same direction: Cursor is becoming an infrastructure company.
Z.ai Ships GLM-5.2 — 1M Context, MIT Licensed — Z.ai launched its new flagship with a 1M-token context window, adjustable reasoning controls, and a coding-first focus. The headline move: it will be fully open-sourced under MIT license, with API and chatbot access following next week. At a moment when Anthropic’s models are literally being shut down by government order, Z.ai is handing its model to anyone who wants it.
DeepSeek Raises $7.4B at $50B+ Valuation — DeepSeek closed the largest AI funding round of the year, with founder Liang Wenfeng personally investing ~$3B. A Chinese government-backed fund invested ~$150M. The deal structure is unusual — investors get economic rights but limited governance power. DeepSeek plans to use the capital to expand compute and advance research. At $50B, it’s now valued at roughly 3x what Anthropic was worth when it raised its last round.
Agentic Code Review: Churn Up 861%, Defect Rate 9% → 54% — The hard part of engineering has moved from writing code to deciding whether to trust it. A Faros AI study found that after agentic coding tools were adopted, code churn increased 861%, per-developer defect rates jumped from 9% to 54%, and review duration increased 441%. The implication: AI-generated code is easy to produce and hard to verify. The review bottleneck may become the defining constraint of the agentic era.
TLDR; (30-sec)
Anthropic suspended Fable 5 and Mythos 5 after a US Commerce Department export-control directive triggered by Amazon researchers who jailbroke safety guardrails
SpaceX is acquiring Cursor for $60B in an all-stock deal — the largest AI developer tool acquisition in history
ChatGPT’s market share slipped below 50% for the first time as users migrate to Gemini, Claude, and Grok
DeepSeek raised $7.4B at a $50B+ valuation, becoming China’s most valuable AI startup
Cursor launched Origin, an agent-native GitHub competitor built for parallel AI agents cloning, branching, and committing at scale
Z.ai shipped GLM-5.2 with 1M-token context under MIT license
OpenAI is preparing GPT-5.6 models for release next week with a 1.5M-token context window
Noam Shazeer, co-author of “Attention Is All You Need,” is joining OpenAI
Anthropic shipped a major Claude Design overhaul with design system imports and code round-trips
A Faros AI study found agentic coding raised code churn 861% and per-developer defect rates from 9% to 54%
💻 Build Tips & Engineering
Securing the Future of AI Agents — Google DeepMind published its AI Control Roadmap: a framework treating internal agents as potentially misaligned, with sandboxing, endpoint security, and prompt injection resistance at the system level. (Jun 19)
Production Infrastructure for AI Agents (Vercel eve) — Vercel open-sourced eve, an agent framework with durable execution, sandboxed compute, approval gates, subagents, and evaluations. Focus on agent behavior; the framework handles production. (Jun 18)
Vercel Connect — Public beta replacing long-lived provider tokens with runtime credential exchange that issues short-lived, task-scoped credentials for agents. Solves the “how do agents authenticate safely” problem. (Jun 18)
NVIDIA Blackwell Sweeps MLPerf Training 6.0 — Blackwell platform dominated MLPerf Training 6.0 with fastest times and largest-scale training at 8,192 GPUs. NVLink and NVFP4 innovations drove the results. (Jun 17)
Never Waste a Token — Practical guide to token efficiency in production LLM systems. Every wasted token is compute, latency, and money — and most apps waste a lot of them. (Jun 17)
Factory 2.0: From Coding Agents to Software Factories — Factory’s software factories are now in production at enterprise scale. Engineers build factories that build software — the role is shifting from writing code to designing the systems that write code. (Jun 16)
🧬 Model Releases
Cursor 1.5T — Cursor/Anysphere announced a 1.5T-parameter model trained from scratch on 100K+ GPUs. Intelligence extending beyond coding, releasing in coming weeks. (Jun 18)
GPT-Bidi-1 — OpenAI’s bidirectional audio model for ChatGPT voice mode. Listens and speaks simultaneously, absorbs interruptions, adjusts mid-sentence. (Jun 17)
VibeThinker-3B — Weibo’s 3B-parameter model posted coding benchmark scores in the same league as Claude Opus 4.5, reigniting the benchmarks-vs-real-performance debate. (Jun 17)
Phi Silica — Microsoft testing small language models engineered to run locally on NPUs of Windows Copilot+ PCs. (Jun 17)
Kimi K2.7 Code — Moonshot’s coding-focused agentic model at 1T MoE parameters. Stronger end-to-end task completion and improved token efficiency vs K2.6. (Jun 15)
🛠️ Tools & Product Updates
Claude Code Now Supports Artifacts — Anthropic shipped artifacts in Claude Code, turning work sessions into live, shareable visual pages with auto-refresh, version history, and privacy controls. Beta for Team and Enterprise. (Jun 19)
OpenAI Prepares GPT-5.6 for Next Week — Mini and Pro variants with 1.5M-token context, improved long-horizon coding, and faster Codex. Competitive pricing aimed at undercutting Anthropic. (Jun 19)
Claude Design Overhaul — Major update: design system imports from GitHub and files, code round-trips, and a fix for the token-burning problem. Prototype toy to enterprise design platform. (Jun 18)
Replit Is Now Available in Claude — Replit integrated directly into Claude, enabling seamless transitions from design to development and deployment. (Jun 18)
Android 17 Expands AI Agent Integration — Android 17 introduced AppFunctions and Android MCP, enabling apps to expose orchestratable tools that on-device agents can discover and execute. (Jun 17)
Sakana Marlin — Sakana’s autonomous research assistant streamlines extensive strategic analysis into hours. Input a topic, Marlin generates a detailed strategy report and slides without human intervention. (Jun 16)
Facebook Gets Its Own AI Mode — Facebook AI Mode transforms the search bar into a conversational tool mining public Groups, Reels, and Marketplace data. Rolling out in the US. (Jun 16)
⚡ Quick Bits
Baseten: $13B AI Startup Betting on Cheaper Alternatives — Specializes in providing software and compute for companies running lower-cost AI models as alternatives to OpenAI and Anthropic. (Jun 19)
DeepSeek Tells Investors “No Poaching” — DeepSeek told potential investors not to poach staff or encourage them to start competing companies — an unusual demand reflecting how talent-constrained the field remains. (Jun 19)
Apple to Raise Prices Due to Memory Chip Crunch — AI infrastructure buildout eating up memory and storage chip supplies. Prices quadrupled since last year, expected to keep rising into 2027. (Jun 18)
Why Is Meta Destroying Its Engineering Organization? — Meta’s move-fast engineering culture demolished in weeks since April. Code churn, forced reorgs, and leadership decisions gutting a proven successful culture. (Jun 17)
OpenAI Leaked Financials: $13B Revenue, $19B Expenses — Revenue grew from $3.7B (2024) to $13.07B (2025). Expenses grew faster, from $7.81B to $19.18B. Hopes to be profitable by 2030. (Jun 17)
Elon Musk Becomes the World’s First Trillionaire — SpaceX shares rose 20% from IPO price. Net worth now exceeds 3% of US GDP — five million times that of a typical American family. (Jun 15)
Meta’s AI Unit Is a “Soul-Crushing Gulag” — Applied AI team on the verge of revolt. A live-streamed presentation was interrupted by an employee meltdown targeting a senior executive. (Jun 15)
📌 Deep Dives
Brain the Size of a Planet: Are LLMs Thonking Too Hard? — Testing Claude 4.6/4.7 and GPT 5.4/5.5 on security vulnerability triaging. Higher reasoning effort and later-generation models aren’t always better — sometimes the overhead makes results worse. A useful corrective to the “more compute = more accuracy” assumption. (Jun 18)
The Once And Future Fable #2 — Zvi Mowshowitz’s second analysis of the Fable shutdown. Key unknowns: what motivated the US government’s decision, how much they understand the underlying technology, whether they’re demanding a narrow fix or a global one, and what they intend next. (Jun 16)
Google DeepMind Explores the Path to ASI — DeepMind published a paper examining how AI could progress beyond AGI toward ASI, identifying four possible pathways, potential bottlenecks, and societal implications. Whether you think ASI is 5 years or 50 years away, this is the most detailed framework for thinking about what the path looks like. (Jun 16)
Inference Cost at Scale with Napkin Math — How to calculate dollar price-per-user from GPU specs, context length, active parameters, and product factors. The back-of-envelope math every AI product manager should be able to do. (Jun 15)
That’s a wrap for this week.
If this was useful, the best thing you can do is share it with someone who’d get value from it too. Forward this email, drop it in your team’s Slack, or just send it to that one friend who’s always asking “what happened in AI this week?” Every share helps us keep this going.
Hit reply if I missed something — I read everything. See you next week.
Stay curious.
Attia