Tokens by Michael Brenndoerfer, Week #1
Anthropic Files Confidential IPO Paperwork for Claude
Today's issue: Claude's fog machine, Copilot's cloud-bill jump scare, and math furniture getting rearranged.
Welcome to #1 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
Main Story
Anthropic confidentially files for a proposed IPO
Anthropic confidentially submitted a draft Form S-1 registration statement to the SEC for a proposed IPO of common stock. In plain English, the Claude maker has started the review process that could let it go public later, but the actual filing details are still private.
The company said any IPO would depend on market conditions and other factors, and it has not set the number of shares or the IPO price. The announcement also uses Rule 135 language, meaning this is a notice of a filing, not an offer or solicitation.
That narrow wording is the point. A confidential S-1 starts the paperwork machine before the prospectus everyone wants to screenshot shows up, which is very corporate and also, regrettably, how adult processes work.
- Draft S-1 filed: Anthropic has started the SEC review path for a potential IPO of common stock.
- Details stay private: No share count or IPO price has been set, so this is process news, not the full prospectus reveal.
- Timing is conditional: Anthropic says any IPO would depend on market conditions and other factors, which is the corporate version of "we'll see how the sprint goes."
- This is not an offer: The Rule 135 language keeps the announcement in notice-only territory, with the real filing still behind the curtain.
Bottom Line: This does not mean Anthropic is immediately going public. It means the company has created the option to do so after SEC review while keeping the important financial details private for now. The signal is procedural: frontier AI labs are starting to look like large infrastructure companies, paperwork stacks included.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
The Feed
-
Microsoft MAI: Microsoft rolled out three homegrown models for transcription, voice, and images, including speech that can generate 60 seconds of audio in one second.
-
Anthropic listed Claude Opus 4.8, an update aimed at better coding, agentic work, professional tasks, and long-running jobs that do not wander off like a bored intern.
-
Copilot billing: GitHub is moving from request units to token-metered AI Credits, so agent-heavy workflows need caps before the bot discovers loop poetry.
-
Devin maker Cognition raised $1B at a $26B post-money valuation and reported $492M run-rate. AI coding agents are now an industrial sport.
-
At Computex, NVIDIA said Vera Rubin is in production, RTX Spark PCs are due this fall, and Vera is aimed at agent workloads. The GPU treadmill remains undefeated.
-
Compute receipt: CNBC says SpaceX's prospectus showed Anthropic agreeing to pay $1.25B a month for Colossus compute through May 2029. That is a spicy invoice.
-
Microsoft caught heat after threatening a security researcher over public exploit disclosures for Defender and BitLocker bugs. Vuln timelines get messy once lawyers, researchers, and public disclosures share a calendar.
-
OpenAI says a reasoning model disproved an Erdős geometry conjecture, with outside mathematicians checking the proof. Math furniture has officially been rearranged.
-
Bumblebee: Perplexity's Go scanner checks packages, MCP servers, VS Code extensions, and browser add-ons as a first-pass AI supply-chain audit.
-
OpenAI listed self-improving tax agents built with Codex, a tidy agent testbed for iterative workflows and a gentle warning to never let my Friday refactor drive.
-
Grok's May sprint added Custom Skills, Grok Build, and connectors for Vercel, Canva, Gamma, and S&P Global. The assistant is getting more app-shaped.
-
Viral image prompts: action figures, Polaroids, scrapbook pages, neo-noir portraits, and clay toys are still the internet's favorite selfie costume rack.
-
Mistral's Devstral 2 123B appeared on the Bedrock tracker at $0.40 in and $2.00 out per 1M tokens, because coding models keep reproducing in the walls.
-
Penn researchers built hybrid light-matter particles aimed at faster, lower-energy AI computing, using physics to bully electrons into better throughput.
-
OpenAI safety stack: a Frontier Governance Framework and third-party evaluation playbook landed, basically rubrics for frontier-model homework before the lab bell rings.
See you next week!
Michael
Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.