Another Wednesday, another selection of AI news and resources to help you become more AI native. This week:

Your vertical SaaS stack is now a plugin.

Whats happening: In one week, Anthropic shipped Claude for Small Business, a back-office bundle with 15 pre-built workflows wired into QuickBooks, PayPal, HubSpot, Canva, and DocuSign. Anthropic also shipped Claude for Legal, a marketplace of agents trained on NDA review and policy drafting. OpenAI launched a Finances tab in ChatGPT that connects through Plaid to 12,000+ bank, brokerage, and credit accounts, with Intuit integration on the roadmap.

Why it matters: AI labs are done shipping horizontal platforms and waiting for customers to build the workflows. They're now packaging the workflows themselves, by industry, and giving them away. The buyer's evaluation just collapsed from "which vertical software" to "does my AI platform have a plugin for this." If your product is a vertical workflow tool, your three-year roadmap is now a competitor's plugin description page.

In the wild: A three-person law firm can skip its legal SaaS and get NDA review from the same tool it uses for email. A small business owner can ask ChatGPT "can I make payroll this month" and get a cash forecast wired to their actual bank balance. Meanwhile Anthropic just landed KPMG's 276,000 employees on Claude. The substitution is no longer theoretical.

Looking ahead: This week, pull your top three workflows and ask honestly: could a generalist AI lab ship a free packaged version of this within 12 months? If yes, that workflow is a feature, not a product. Deepen the moats a plugin can't replicate (proprietary data, regulated integrations, network effects), or get acquisitive while your customer base is still defensible.

Zapier just open-sourced AutomationBench, a free benchmark that measures how cheap AI models actually perform on real production workflows.

Wade Foster's team published it after finding a 100x to 1800x cost spread between expensive and cheap models on the same tasks, with the cheap ones handling the work fine.

Pro tip: Anything you run more than 100 times a day is where the gap bites. Ten thousand tasks at $1.80 each is $18,000. At $0.02 it's $200.

Click here to read the full article

Google made Gemini the layer beneath every Google app at I/O 2026.

At Tuesday's I/O, Google rolled out Gemini 3.5 plus Spark (a 24/7 personal agent with local computer access this summer), Omni (a model that generates video from any input), Antigravity 2.0 (parallel coding agents on a managed desktop), and AI Studio that builds full Android apps from a prompt. Search, Gmail, Docs, YouTube, Android, and the developer stack now route through Gemini. The pitch shifted from best chatbot to operating layer underneath every Google product.

LINK

Andrej Karpathy joined Anthropic to run a new pretraining unit.

OpenAI cofounder and former Tesla AI lead Andrej Karpathy is joining Anthropic's pretraining team, the group behind Claude's core capabilities. He'll lead a new unit focused on using Claude to accelerate pretraining research itself. The announcement crossed 17M views on X in 24 hours. The non-AI translation: Ronaldo just signed for Manchester United.

LINK

A California jury rejected Musk's OpenAI lawsuit in under two hours.

After a three-week trial featuring corporate emails, personal diary entries, and Mira Murati testimony, the 9-person jury unanimously ruled the statute of limitations had expired on Musk's claims that Altman "stole a charity." Musk plans to appeal. The decision clears a major obstacle ahead of OpenAI's expected IPO later this year, with the company now valued at roughly $852B. Musk had invested ~$38M during OpenAI's early nonprofit years.

LINK