Welcome back! Reco just raised $55 million to help companies secure their AI agents, which tells you a lot about where the market is headed. Enterprises are apparently deploying agents so fast that security teams are now discovering thousands they did not know existed. Incredible news for anyone who thought “shadow IT” was not stressful enough. We have officially entered the era of needing software to find the software pretending to be your coworkers.
Did Google Take the AI Lead?
Another day, another "best model in the world"… and this one comes from a place I don't think most people saw it coming from.
On Tuesday, Google DeepMind unveiled Gemini 4 Argon, its new frontier model—and by the numbers, it's a legitimate contender for the strongest AI model on the market right now.
What stands out:
1 million token OUTPUT limit, up from 64K. For context, GPT-6 Astra's max output is 128K. Argon can generate hundreds of thousands of tokens in a single pass, opening the door to actual long-horizon reasoning instead of chopping big problems into smaller chunks.
Trained for cybersecurity defense. Argon can autonomously find, validate, and patch critical software vulnerabilities. Google says it already found a critical flaw in healthcare software used by hospitals worldwide—one that earlier frontier models had missed.
Industry-leading resistance to indirect prompt injection, according to Gray Swan's benchmark.
Built for enterprise-grade knowledge work like legal, finance, software engineering, long multi-step research.
The catch? You can't actually use it yet.
Argon is rolling out first through Google DeepMind's Fairwind Program, a trusted-tester cohort of cybersecurity defenders who get access to the model with its cyber guardrails turned off. Public API access, pricing, and consumer availability all have no date attached.
My POV: It's another entry in a trend we've been tracking all year: frontier models launching to a small group of trusted testers before any public access. We saw it with Anthropic's Mythos 5 and Fable 5. We're seeing it with OpenAI's "Critical cybersecurity capability" tier on GPT-6. And now Google's doing it with Argon. The "release to everyone on day one" era seems to be ending at the frontier.
The question that keeps nagging me: if the most powerful AI capabilities are increasingly concentrated inside a small group of approved institutions, what happens to everyone else?
Hit reply and let me know: do you think Google took the lead?
— Matt
OpenAI Launches Cheaper Sol and New Dots Agents
Via AP
OpenAI launched GPT-6.1 Sol and introduced Dots this week, giving users a cheaper model and a new personal agent product.
The agent push: Dots is OpenAI’s answer to Meta’s Muse.
The agents are always-on assistants that can work across connected apps, learn user preferences over time, and operate through a cloud computer with access to a browser and more than 4,000 supported apps.
Users can interact with Dots through a text-message-like interface, voice calls inside ChatGPT, and integrations with Slack and Microsoft Teams.
The cheaper model race: GPT-6.1 Sol is OpenAI’s new lower-cost model for agentic coding, computer use, and professional work.
OpenAI says it gets close to GPT-6 Astra’s performance while costing one-fifth of the standard input and output token prices. It is also meant to improve on GPT-6 Sol across programming, debugging, document understanding, and multistep workflows.
Sol’s positioning looks a lot like what Anthropic is doing with Sonnet: not the absolute top model, but a faster, cheaper one that can handle the work people run all day.
The bigger picture: AI labs are learning that users do not just want the smartest model in a vacuum. They want something convenient, fast, reliable, and cheap enough to use constantly. Dots is OpenAI’s bet on convenience and stickiness. GPT-6.1 Sol is its bet on cost and speed. Together, they show where the next phase of the market is going: agents that do more, powered by models that cost less to run.
Tech Leaders Sign AI Safety Deal
Top AI executives signed a new voluntary safety agreement with the Trump administration, laying out how major labs plan to self-police frontier AI.
What’s in the deal: Leaders from Google, Anthropic, Meta, OpenAI, xAI, and Nvidia signed the agreement, called the Joint Commitment on Frontier Responsibilities. The companies are committing to four broad safety steps: internal monitoring of model capabilities and risks, dedicated internal teams responsible for compliance, independent external audits or evaluations, and board-level oversight through an independent committee.
What it doesn’t do: The agreement is “morally binding,” but it does not appear to include enforceable penalties. That means the deal is more of a public commitment than a regulatory regime. It gives the administration and the companies a way to say they are taking safety seriously without creating a hard legal framework around frontier model development.
Why now: The deal comes after months of louder AI safety warnings, including rogue agent incidents, model-release delays, and growing concerns that labs may be building systems faster than they can reliably monitor or control them. The timing also matters politically: the Trump administration is trying to preserve a light-touch approach while still responding to fears that powerful AI systems could cause real-world harm.
The bigger picture: Regulators are looking for signs that labs can govern themselves. Labs are looking for a way to keep building without triggering harder rules. The question is whether voluntary commitments, audits, and board committees are enough when the same companies promising restraint are also racing to win the next model cycle.
Hire a Team of AI Agent Coworkers
Imagine opening Slack at 8 a.m. to find a competitor briefing already waiting. Nobody prompted it. Your AI agent proactively made it. And used other AI agents to help it. That reality exists with Hyperagent.
Hyperagent lets you hire a team of specialized agents that can:
Take on distinct jobs, like finance analysis or marketing research.
Hand work to one another instead of leaving you to coordinate every step.
Serve your whole team through shared agents, rather than living in one person’s setup.
Work within clear boundaries and approval rules for each role.
Want a fleet of AI agents your whole team can actually use? Sign up for a paid plan and get $100 in bonus credits.
Edit Videos With AI
Via Vidmoat
Vidmoat is a browser-based AI video editor that takes a text prompt, plans edits, and self-corrects until the task is complete.
How you can use it
Cut dead air from videos automatically
Generate synced captions
Remove video backgrounds
Export clips for TikTok, Reels, YouTube, and Shorts
Pricing: Free and paid
Deploy Secure AI Agents
Via Globster
Globster is a platform for deploying secure AI agents inside organizations. Each agent runs in an isolated sandbox with a dedicated credential vault and configurable human approval controls, while teams can connect thousands of apps and interact with agents through Slack, WhatsApp, or Telegram.
How you can use it
Deploy agents in isolated sandboxes
Store credentials in dedicated vaults
Add human approval controls
Connect agents to 3,000+ apps
Pricing: Paid
Jobs, announcements, and big ideas
Anthropic targets mid-November for its IPO.
Amazon pledges to invest $1B over the next five years in communities where it operates data centers.
OpenAI fires three researchers who violated company policies on sensitive information.
Reddit shuts down RSS feeds because of AI bot scraping.
SpaceX launches Google AI chips into orbit.
Shopify debuts Canvas, an AI-powered website-building tool for merchants.
Dev day, new models, tons of announcements. All the AI news you need to know from this week.
That’s a wrap! See you next week for more.
