Anthropic Aims at SpaceX's Record IPO. Codex Just Opened iMessage.
Yesterday covered teen defaults, a Marvell TPU-attach warrant, and a laptop SWE-agent bench. Today the tape splits: a Bloomberg-sourced Anthropic IPO size target, and OpenAI’s Aug 20 Codex changelog.
Capital is trying to print a Claude listing that matches a rocket company. The coding agent on your Mac just asked to read your texts.
Lead: Anthropic wants SpaceX’s raise, not SpaceX’s ticker
Bloomberg, via The Straits Times (Aug 21): people familiar with the matter say Anthropic expects to match or beat SpaceX’s record first-time share sale. SpaceX raised 75 billion at the outset; the overallotment took the final figure to 86.2 billion. Anthropic is running the numbers ahead of a public filing as soon as the end of August. CFO Krishna Rao’s recent investor briefings, the same people said, skirted valuation.
This is raise size, not a 2 trillion market-cap promise. A first-time sale that tops SpaceX would push 2026 US IPO volume past 2021’s 195.2 billion watermark — year-to-date through Aug 19 was already 160.6 billion.
The operating tape behind the rumor is the part that matters for anyone buying Claude tokens:
| Number | What it is | Source frame |
|---|---|---|
| 965 billion | May 2026 private valuation after a 65 billion raise | Straits Times / Bloomberg |
| 852 billion | OpenAI’s March valuation on a 122 billion raise | same |
| >11.5 billion | Preliminary Q2 revenue vs 787 million a year earlier | same |
| 65 billion | Annualized run rate by end of July | same |
| ~42 billion | 2025 net loss (about 5× 2024’s ~8.3 billion) | same |
| >10 billion | Revolving credit facility being finalized | same |
| ~2% | Dario Amodei’s reported stake, with super-voting shares under discussion | The Information via Bloomberg |
AFR’s Bloomberg reprint is the same story: Morgan Stanley, Goldman Sachs, and JPMorgan on the book; other banks may join. Anthropic is on track to list ahead of OpenAI, which Bloomberg has pegged at a 2027 window. Both have already filed confidentially.
Bold insight: the P&L is the product. Positive adjusted operating income in Q2, a 65 billion run rate, and a 42 billion full-year loss in 2025 can all be true at once. That is what a frontier lab looks like when compute contracts (including a SpaceX deal that sources said could be tens of billions over three years) hit the income statement before the models amortize.
None of this is an S-1. Details can change. Treat “match SpaceX” as a target the bankers are modeling, not a priced deal. Super-voting stock for a CEO at ~2% is the governance tell: public capital, private control. If you are an operator, the question is not “is Claude a public company yet.” It is whether a 65 billion run-rate lab still prices API the way a private lab does once quarterly guidance exists.
If you are actually filing models, not watching IPOs: a USB-C dock and a second monitor still beat a Bloomberg terminal for the work that pays for Claude.
Secondary: Codex can read iMessage. Approval is the product.
OpenAI’s Aug 20 release notes shipped the agent-surface story yesterday’s post did not cover. On Apple silicon Macs, the Apple Messages plugin in the ChatGPT desktop app can read and search iMessage, SMS, and RCS and prepare or send through Messages. Default: ChatGPT asks you to approve the message and the recipients before it sends.
9to5Mac is the operator page. It is all plans, ChatGPT Work and Codex — not regular ChatGPT chats. It does not let you talk to ChatGPT remotely through Messages. Intel Macs are out. OpenAI’s plugin guide flags persistent-approval risks, revocation steps, and a known issue with tasks that disable approval prompts.
Same changelog, three more surfaces:
| Ship | Who gets it | The catch |
|---|---|---|
| Apple Messages plugin | All plans, Apple silicon Mac, Work + Codex | Default approve-to-send; persistent approval is the footgun |
| Computer History | Pro users, now EEA, Switzerland, UK | Off by default, requires Memories; interaction events, not screenshots |
| Codex read-only snapshots | Personal share links (anyone with the URL) | Static; no tool calls or shell I/O; redacts known secrets — still review paths, diffs, images |
| Pinned chat sync | Desktop app ↔ iOS, same Codex account | Android not in this drop |
| ChatGPT Site URL change | Plus/Pro site owners | Old URL redirects, custom domains unchanged |
Bold insight: this is the same subtractive-defaults week as teen auto-enroll. The coding agent is no longer sandboxed to the repo. It is sandboxed to whatever local surface you attach, with an approval checkbox that tasks can apparently turn off. A snapshot that strips shell I/O is the honest admission: the interesting leak is never the export AWS_SECRET. It is the path, the diff, and the screenshot you forgot was in the thread.
If you turn this on, treat Messages like prod credentials. Revoke when you are done. A privacy screen is theater next to “Codex may send without a prompt if you enabled persistent approval.” An external SSD for local clones still beats letting the agent rummage iMessage for a ticket number.
Takeaway
- Today: the Anthropic story is a raise-size target and a month-end filing window, sourced to people familiar — not an S-1. Watch the credit facility and super-voting language, not the 75 billion headline.
- On the Mac: if Codex can see Messages, decide persistent approval on purpose. The known-issue note about tasks disabling prompts is the actual changelog.
- Snapshots: review before you paste a Codex share URL. Redaction is pattern-based. Paths survive.
The spaghetti this morning is a lab that wants a rocket-company IPO and a coding agent that wants your iMessage history. One of those needs an S-1. The other needs a checkbox you should not leave on.
Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.
Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.