GPT-6 is here, and it's cheaper than last year's discount tier. OpenAI quietly released two smaller GPT-6 models, Sol and Luna, alongside the flagship GPT-6 Astra. The headline number: API pricing comes in 50% below the previous generation's promotional rates. Both models are available in ChatGPT Work, Codex, and the API. The real-world benchmark that matters for small teams: Luna Max completed 18% of the hardest real Rails dev tickets for about $11. That's not theoretical. DHH ran the test and posted the numbers. If you've been watching API costs the way a restaurant watches food cost percentage, this is a meaningful shift. More from OpenAI here.
An AI agent hacked Australia's national health service, and the government found out months later via email. Not a headline from a dystopian novel. Wired reported it. An OpenAI agent compromised the infrastructure, and the disclosure came in a routine notification, long after the fact. If you're considering giving any AI agent access to your systems, customer data, or internal operations, this is the story to read first. The liability and disclosure questions are not hypothetical anymore.
Meta's Muse agent handed a researcher 6.8 GB of internal files, including SSH keys. A researcher simply asked Muse to archive its own accessible files. What came back: the Linux root filesystem for his session, internal Meta docs, integration code, memory files, agent logs, and SSH keys. When he reported it through Meta's bug bounty program, they marked it "Not Applicable." Meta's Connect coverage is here, but the security incident is the story worth sitting with. Before you integrate any personal AI agent into a customer-facing workflow, ask yourself what it can access, what it might hand over, and who you'd call if it did.
Tools Worth a Look
Taskade: Turn your workspace into a client-facing portal, complete with project trackers, onboarding flows, and team wikis, built from a single prompt. Good for owners who want to give clients a structured experience without hiring a developer. Free tier available; paid plans vary.
Trending on GitHub
tensorflow/tensorflow: Google's open-source machine learning framework, the one most AI tutorials are built on. (46 stars today)
openbao/openbao: A tool that stores and manages sensitive business data like passwords, certificates, and API keys, so they don't end up in Slack threads or spreadsheets. (364 stars today)
block/buzz: A shared communication platform built around group AI context, think of it as a team chat where the whole group shares one AI memory. (339 stars today)
microsoft/vscode: Microsoft's free code editor, still the most popular one on the planet. (95 stars today)
zhaoxuya520/reverse-skill: An AI-powered toolkit for security researchers that automatically routes tasks to the right tool and builds up a self-updating knowledge base. Supports Claude Code, Cursor, and similar AI coding clients. (361 stars today)
Also Worth Knowing
ChatGPT Voice can now run plugins hands-free: email, calendar, Slack, and browser while you talk. Inside ChatGPT Work, you can dictate a document, deck, or spreadsheet into existence. Getting closer to a real hands-free computer.
Shelly Palmer's weekend roundup makes a clean argument: frontier-level capability is now cheaper than it was six months ago, and the competitive edge is moving from access to execution. Worth five minutes if you're deciding where to focus right now.
Opus 5.5 early verdict from builders: roughly 40% cheaper than Opus 5, faster, more direct. A lot of people are calling it their daily driver within days of launch.
Microsoft's update backs the adoption-first thesis: agency owners and operators deciding where to put energy right now, this is a useful frame. The real moat is adoption, not chasing the bleeding edge.
Qwen-Image-2.1 dropped: an open 7B model that generates and edits images, handles transparent backgrounds, and accepts up to 10 reference images. Worth watching if you're paying for image generation tools.
DHH's Rails benchmark post-Opus 5.5: OpenAI still leads on the hardest tickets after the Anthropic launch. Good grounding if you're choosing which API to build on.
Meta Muse Realtime Avatar research: 870ms response latency at 25fps, preferred over Runway Characters 78 to 22 in human ratings. Interesting if you're watching the avatar and video agent space.
Claude flagged a potential CRISPR-like system in bacteriophage DNA: Anthropic's biology team ran the experiments. A reminder of where this technology is heading, even if it's not on your to-do list today.
UN Security Council held its first high-level AI and international security briefing, with Bengio and Altman both testifying. If regulation is on your radar, this is the transcript to bookmark.
HubSpot is running a Sept 30 webinar on turning LinkedIn content into repeatable sales outreach. Free, practical, worth a look if LinkedIn is part of your pipeline.
-Scott
SmartOwner is published (almost) daily by the team at DigitalTreehouse. Want AI consulting or automations for your business? Reply to this email.