A hundred billion and counting
Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt.
WHAT'S HOT
💰 Anthropic's annualized revenue is on track to pass $100 billion. The run rate went from $9 billion at the end of 2025 to $65 billion in July to this, and the IPO is now targeted for November at a valuation around $2 trillion, which would raise more than SpaceX did in June. (Bloomberg)
🧠 Claude Opus 5.5 landed this morning, 20% cheaper than Opus 5. $4 per million input tokens and $20 out, with Anthropic claiming Fable 5.1-level performance on most tasks and 40% lower running costs on typical workloads. Sonnet and Haiku versions follow in the coming weeks. (TechCrunch)
🔓 One zero-click exploit hit every major coding agent at once. Plugin4Shell lets a malicious plugin update run code with your agent's permissions, no click required, because the agent checks out the pinned commit and never verifies it landed there. Claude Code is patched in 2.1.179 and Codex in 0.146.0. Copilot CLI has no fix yet and Gemini CLI isn't getting one. (Help Net Security)
⚖️ Universal and Sony are suing Suno again, over the licensed model. Their Boston complaint names 60,202 recordings and argues v6 inherited the infringement because it was trained on outputs of the earlier models, "the fruit of the same poisoned tree." Warner, BMG and Believe licensed v6. These two didn't. (Music Business Worldwide)
🏛️ California is studying a kill switch for frontier models. Newsom's executive order gives the Government Operations Agency until November 16 to recommend whether the law should require one, whether independent auditors should sit inside the biggest labs, and what else counts as a reportable safety incident. (Governor of California)
48 •
Productivity
Nobody reads the output anymore
A model lands about every week now, so Kristina Grits (@kristina__grits) asked whether people actually test them as they arrive. Aleksandar Blazhev (@byalexai) tests almost every drop, because he thinks older models get worse once a new one ships and he won't get attached to any of them.
Siarhei Hamanovich (@siarheihamanovich) doesn't bother on release day, since swapping models is one line for him. Working out whether the new one is better is the expensive bit, and his team never reads the output to decide: "reading the output always makes the new model look better." They run fixed graders instead. Aryaman Sharma (@aryaman_sharma2) called the alternative productive-looking procrastination.
Good thread if you upgraded a model last week and still can't say whether it helped.
```
