Cursor's new owner is Musk's AI company — a $60 billion deal

SpaceXAI closes a $60 billion all-stock deal for Cursor's maker. Plus: Stripe reportedly near a $7B OpenRouter buyout, Claude's system prompts published, Gruber vs. the watermark, a 232x Codex speedup with a catch, and OpenAI's Astra pause.

Share
Cursor's new owner is Musk's AI company — a $60 billion deal

Cursor spent the weekend changing hands. SpaceXAI, the Elon Musk AI company built from what used to be xAI, closed a $60 billion all-stock deal for Cursor's maker, Anysphere, folding one of the most-used coding agents into the same stack as Grok. Stripe is reportedly circling a $7 billion buyout of OpenRouter, the router a large share of coding agents use to switch between models — a second piece of coding-tool infrastructure changing hands in one week. Anthropic published the actual system prompts running Claude's web and mobile apps, no NDA required, and Daring Fireball's John Gruber spent the same week arguing Claude's new text watermark makes Claude worse at the one job it has, writing. And a solo developer's claimed 232x GPU-kernel speedup, from a loop driving OpenAI's Codex agent, drew skeptics on Hacker News. Plus a catch-up this newsletter owed you: OpenAI's next model, Astra, and why its coding ability is the stated reason parts of its development are paused.


Who owns the tools now

SpaceXAI buys Cursor for $60 billion — the coding tool joins Musk's stack

Elon Musk's xAI rebranded as SpaceXAI earlier this year. On August 14 it closed an all-stock acquisition of Anysphere, the maker of Cursor, at a $60 billion valuation, the largest startup acquisition on record, per an SEC Form 8-K. Cursor's own blog confirms the deal but names no financial terms. Those come from the filing and four financial outlets reporting roughly 391 million SpaceX shares issued. Cursor now runs as a SpaceXAI subsidiary with access to its Colossus compute cluster, and Cursor's post cites Grok 4.6 as an early result of the combination. One of the most-used AI coding agents changed owners over a weekend.

Cursor is now a part of SpaceX · Cursor
Cursor has officially been acquired by SpaceX.

Stripe reportedly nears a $7 billion deal for OpenRouter

Bloomberg reports Stripe has agreed to buy OpenRouter, the service that lets a coding agent call GPT, Claude, or Gemini through one API and switch providers without rewriting code, for over $7 billion. Neither company confirms it. Stripe told TechCrunch it doesn't comment on rumors. OpenRouter's last disclosed valuation was $1.3 billion, in May. If it closes, the company payments infrastructure runs through would run the plumbing a large share of AI coding tools use to reach their models.

Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+ | TechCrunch
OpenRouter’s CEO recently described the startup as Stripe for AI.
Why these two run together. Two pieces of AI-coding infrastructure changed hands, or came close to it, in the same week: one deal closed (SpaceXAI/Cursor), one is still a rumor (Stripe/OpenRouter). Neither is a model release. Both are about who controls the layer underneath the tools this newsletter covers.

What Claude tells you about itself

Anthropic publishes Claude's actual system prompts: six model versions, diffed

Anthropic put the real system prompts for every recent Claude web and app version on a public docs page, across six model versions, each diffed against the last. The prompts cover the mundane (today's date, so Claude doesn't guess) and the consequential (child-safety rules, how hard to push back on a bad idea). None of this reaches the API. Only claude.ai and the apps carry a system prompt. Anyone building on Claude can read the actual instructions, not guess.

System Prompts
See updates to the core system prompts on [claude.ai](https://claude.ai) and the [Claude iOS app](https://anthropic.com/ios) and [Claude Android app](https://anthropic.com/android).

Gruber: Claude's watermark is "a perversion of writing"

Daring Fireball author John Gruber objects to Claude's text watermark (disclosed August 11). Anthropic said the watermark wouldn't "change meaning, quality, or readability," but the mechanism nudges word choice probabilistically toward a secret "green" list and away from a "red" one, which is exactly a quality change. Gruber's line: "The idea that anything other than my needs should factor into the generation of text for me is patently offensive." He names no specific ruined sentence. His objection is to the premise that synonyms are interchangeable.

Anthropic’s ‘Watermark’ Text Adulteration in Claude Is a Perversion of Writing
It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of embedding hidden clues within the text to suggest its provenance. The idea that anything other than *my* needs should factor into the generation of text *for me* is patently offensive.
Why these two run together. Anthropic spent the same week publishing exactly what it tells Claude to do (item 3) while a critic argued its public claim about a different instruction — that the watermark doesn't touch quality — doesn't hold up. Anthropic published what it tells Claude to do. Gruber is testing whether one of its public claims holds.

Shipping with AI: what works, what breaks

A Codex loop hits a 232x GPU speedup — the contest's top entries overfit

Sankalp placed 12th of 183 in GPU Mode's kernel-optimization contest, reporting a 232x speedup over the torch.geqrf baseline: 1,500 leaderboard submissions in 14 days, OpenAI's Codex running under /goal, and a beam of 3-5 candidates, not one champion. The catch, per an HN commenter: 8 of the top 10 entries broke on inputs outside the competition's, and the author notes the fastest kernels exploited the test distributions. Read 232x as contest-specific, not general.

Auto-research with codex: How I achieved a 232x Faster Kernel over baseline with Codex in GPU Mode’s qr_v2 problem
Table of Contents Intro Contest in short Problem intro Why this problem is auto-research-able Learning Enough to Ask Better Questions (Optional) Math f…

What OpenAI is holding back

OpenAI pauses parts of its Astra work, citing coding gains

OpenAI named Astra, its next major model, on August 1 by publishing ten machine-checked proofs of open math problems. Six days later it said internal evaluations "indicate significant advancements in agentic coding and cybersecurity," that it "cannot rule out critical cyber capabilities," and that it is pausing internal Astra activities that don't meet its strengthened security controls. No release date exists in OpenAI's statements. The circulating dates come from prediction markets and SEO roundups.


Also worth your time

  • Grok 4.6 landed in GitHub Copilot on August 14 — the same model Cursor's new owner just cited as an early result of the SpaceXAI/Cursor combination is now also available in a second major coding tool, per GitHub's changelog.
  • A Claude Code diagram-generation skill is spiking on GitHubcathrynlavery/diagram-design gained roughly 15,600 of its 20,200 stars this week. No launch post found, and no independent hands-on beyond the star count itself; watching, not running, until there's something to verify against.

Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.

The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.