Good morning. The week ends on a deal that nobody had on their bingo card: SpaceX is buying Cursor for $60 billion in stock, days after its IPO and amid the wreckage of xAI. Elsewhere, ChatGPT’s grip on the assistant market finally cracked below half, Anthropic is somehow winning business while the White House yells at it, and the Fable saga picked up a new wrinkle around what actually triggered the ban.
SpaceX buys Cursor for $60B. Per Reuters and TechCrunch, the all-stock deal folds Cursor into SpaceX’s post-IPO AI rebuild, which has been a mess since the xAI merger — all 11 co-founders have left, Grok keeps embarrassing itself, and Musk has publicly conceded the company “was not built right.” The price is roughly 24x what Microsoft paid for Minecraft, for a company that wasn’t on track to break even despite raising over $3B. HN’s reaction was a mix of disbelief and indifference, with one commenter capturing the mood: “Is anyone on HN still actually using Cursor in 2026?” Most said they’ve moved to Claude Code or Codex.
ChatGPT slips below 50%. ChatGPT’s global market share dropped to 46.4% in May, per TechCrunch, with Gemini at 27.7% and Claude at 10.3%. ChatGPT still has 1.1 billion monthly users, but the assistant market is maturing: growth rates are decelerating, users are switching more readily (OpenAI’s Pentagon deal caused a measurable uninstall spike), and Claude is monetizing harder, with 13% of users on paid tiers.
Anthropic’s “dangerous” label is selling. Despite — or because of — the export ban, Ramp data shows Anthropic overtook OpenAI in business AI subscription share for the first time in May, 41% to 39.5%. Ramp’s lead economist pointed out that Anthropic’s best month for business adoption coincided with the Defense Department labeling it a supply-chain risk. Being declared too dangerous for the government, apparently, makes for great enterprise marketing.
The Fable ban was triggered by “fix this code.” Adding to what we covered earlier this week, The Register reports — via Katie Moussouris, the only outside expert to read the underlying report — that the “jailbreak” was researchers asking Fable 5 to fix vulnerable code, then writing test cases to confirm it worked. No guardrail bypass, just the model doing defensive security work. Wired’s parallel piece makes the obvious follow-up point: similar capabilities are coming from OpenAI and others within months, so the export controls are a delay, not a fix.
Local models cross a threshold. Vicki Boykis argues that Gemma 4 and Qwen 3 now hit roughly 75% of frontier accuracy on agentic coding tasks running on a 64GB M2 Mac — Python refactoring, unit tests, bootstrapping ML repos. HN was split: enthusiasts pointed to Qwen3.6 27B and Gemma variants as genuinely usable, while skeptics noted the dense-vs-MoE tradeoff is still painful (dense models are smarter but slow, MoE is fast but unreliable). One commenter switching from Qwen3.6 back to Sonnet 4.6 called it “such a downgrade,” which is either a sign of the times or a sign they need new hardware.
Qwen ships a robotics suite. Alibaba’s Qwen team released the Qwen-Robot Suite, a foundation model stack for embodied AI. HN’s most interesting thread was strategic rather than technical: one commenter noted the TAM for robotics dwarfs coding or services, and Qwen’s manufacturing ties could put a million units a year into the field within three years. Whether the models actually solve real-time world-state prediction — say, catching a ball — is unanswered.
The Netherlands tries sovereign AI on €13.5M. TNO, SURF, and the Netherlands Forensic Institute are building GPT-NL from scratch on lawfully sourced data with open-source code. HN commenters were unconvinced the budget is remotely enough — one pointed to Sweden’s GPT-SW3 as a cautionary tale — and argued national labs should fine-tune Qwen or Kimi instead of starting from zero. Dutch tech press is reportedly already souring on the project.
That’s the week. Cursor finds a strange new home, Anthropic turns a regulatory beating into a sales pitch, and local models inch closer to making any of this matter less. See you Monday.