The AI Hub Weekly: ChatGPT Agents Grow Up, GPT-5.6 Rewrites The Rules

A week that started with an upgrade to the models powering everyday tools turned into a turning point for “AI agents” — the kinds of assistants that actually do things for you, not just talk. OpenAI and Google both made AI feel busier, smarter, and far more hands-on, and: you can try these changes right now.

TOP STORIES

OpenAI Puts Supercharged GPT-5.6 Into Microsoft 365 Apps

• OpenAI’s new GPT-5.6 language model is now the heart of Microsoft 365 Copilot, powering Word, Excel, PowerPoint, and more.
• This means faster, more accurate AI assistance is coming to millions — no extra steps or upgrades needed if you already use Microsoft 365.
• GPT-5.6 promises better writing, analysis, automation, and understanding of complex instructions than before.
• Source

ChatGPT Work: Your AI Can Now Manage Projects — Not Just Answer Questions

• OpenAI launched ChatGPT Work, a “work agent” that can stay on a project for hours, act across your apps and files, and deliver finished work like spreadsheets or slides.
• Unlike standard chatbots, this agent remembers project goals, pulls info from connected apps, and helps organize longer workflows.
• Available to ChatGPT paid users starting this week, but full integrations may require permissions or supported services.
• Source

GPT-Live Brings Real-Time, Two-Way Conversation to AI Voice

• OpenAI unveiled GPT-Live, an upgrade that lets AI talk with you in real-time, listening and responding at the same time — more like a real person.
• GPT-Live can interrupt, use conversation fillers (“mhmm”, “yeah”), adapt to your speaking, or stay silent when asked.
• Available now to select ChatGPT users, with wider rollout expected over the next few months.
• Source

OpenAI Rolls Out GPT-5.6: New Flagship Models in Three Tiers for Any Task

• The GPT‑5.6 family is officially available: Sol for top intelligence and efficiency, Terra for balanced everyday use, and Luna as a cost-saving option.
• These models boost coding, research, writing, and even cybersecurity — with claims they hit state-of-the-art benchmarks.
• Pricing and access depend on which model and platform you use (API, ChatGPT, Copilot, etc.).
• Source

OpenAI Expands Bio Bug Bounty to an Ongoing, Private Security Program

• Their biosecurity “bug bounty” (rewards for finding flaws or risks in Bio-related AI tasks) is now a permanent, private effort open to select safety experts.
• The program tests for universal AI jailbreaks in biological domains, hoping to prevent future misuse or accidents with advanced AI models.
• If you’re a security or biosafety researcher, you can apply; regular users just benefit indirectly from tighter protections.
• Source

New Guide Makes Getting Started With ChatGPT Easier Than Ever

• OpenAI published a beginner tutorial on how to use ChatGPT, demystifying what it can (and can’t) do for people new to the tool.
• The guide explains how to start a conversation, what the AI is good for, and offers tips for better results — no jargon required.
• Useful for anyone overwhelmed by all the new features or curious about jumping in for the first time.
• Source

THIS WEEK IN AI

What stood out this week wasn’t just the relentless pace of model upgrades — it’s how quickly the “AI agent” has become reality for the rest of us. If you think of chatbots as simple answer machines, throw that image out. With ChatGPT Work and upgrades like GPT-Live, we’re being asked to trust AI not just to research, but to plan, coordinate, and execute real tasks end-to-end.

OpenAI’s not alone here: Google and others are racing to make agents that can actually move through your files, email, calendars, and productivity apps, noticing what you want and pro-actively turning loose goals into results. The stakes are higher than “write me an email.” In theory, you can now ask your AI to launch a project, coordinate files, and update your workflow — things humans used to be essential for. We’re all guinea pigs in a massive experiment on digital delegation.

But here’s the tension: as the tech gets more autonomous, it also gets more opaque. What does it actually mean to “assign” a goal to an agent for hours? If it links to your accounts and acts on your behalf, do you trust it to stay on track, or to remember what’s private? These features demand you get comfortable with oversight, permissions, and a certain amount of trial-and-error — and tech companies promise it’ll all feel natural, but that’s far from guaranteed.

And look at the simultaneous emphasis on security, like the expanded bio bug bounty, and the push for training about “how to actually use ChatGPT.” We’re past the phase where AI is exciting just for its raw power. Now it’s about who gets it to work for them — and how safely. Tools are only as good as the control and context we have.

So: Are you brave enough to hand off actual work to an AI agent — while keeping watch? Try out the new ChatGPT workflows, poke at the permissions, and see where it helps (or breaks). The new boundaries are being drawn in real-time, and users willing to experiment will shape what “personal AI” really means.

MORE TOP STORIES

GPT-5.6 Models Now Ready for All: Choose From Sol, Terra, or Luna

• The latest generation of OpenAI’s language models (Sol, Terra, Luna) is now generally available across multiple products and APIs.
• Each option is optimized for a different balance of speed, intelligence, and cost — with Sol being the premium tier.
• You may need to choose your preferred model in some developer settings; mainstream apps like ChatGPT pick defaults for you.
• Source

Japan’s MUFG Rolls Out ChatGPT Enterprise to 35,000 Employees

• Mitsubishi UFJ Bank, Japan’s biggest, is now running ChatGPT Enterprise across its 35,000-strong workforce.
• The aim: streamline back-office operations, create better customer experiences, and make employees more productive using AI-powered chat.
• Could set a precedent for mass-scale enterprise AI deployments outside tech-heavy sectors.
• Source

Australian Payments Plus Says ChatGPT and Codex Are Saving Staff Hours Every Week

• Australian Payments Plus (AP+) says most employees using ChatGPT Enterprise and Codex are saving 2+ hours per week.
• 80% of surveyed staff say creativity and work quality improved, and simulation building time dropped from days to a single day.
• Real productivity boosts are showing up, not just in theory but in workers’ own feedback.
• Source

Google’s Gemini API Adds Background Tasks and Smarter Agent Features

• Google updated its Gemini API (used by developers to build with AI agents) to allow agents to run tasks in the background and better handle user permissions and credentials.
• New features target reliability and real-world usage — not just lab benchmarks.
• If you use apps powered by Gemini, expect smoother agent-driven experiences rolling out soon.
• Source

Hugging Face and NVIDIA Release Massive New Dataset for Building AI Agents

• The open “Data for Agents” initiative just shipped richer training data, including code, workflow traces, and troubleshooting logs to help developers build more robust AI agents.
• The aim: get AIs to better recover from “weird” situations, such as failed API calls or new workflows.
• End users won’t interact directly but will benefit as future AI agents become less likely to crash or flounder when things don’t go as planned.
• Source

Hugging Face Adds Blazing-Fast Backend to Open-Source AI Transformers

• The “vLLM” backend now lets open-source AI models (like Hugging Face Transformers) run at near-native speeds — much closer to proprietary AI from tech giants.
• More speed means cheaper, more responsive AI tools for companies and hobbyists who want open alternatives to OpenAI or Anthropic.
• Update is live now for those maintaining or running their own AI-powered software.
• Source

ALSO THIS WEEK

• AI agent crawlers now need permission — Cloudflare will begin blocking “AI agent” bots from accessing parts of the web by default starting September 15; site owners must now opt in. (Source)
• Directly Responsible Individuals (DRI) — A primer on what “Directly Responsible Individuals” means in tech orgs. (Source)
• Fable gets another bump — Anthropic once again postpones retirement of their Fable AI model in Claude Max plans. (Source)
• OpenAI’s Head of Safety Is Leaving the Company — OpenAI’s top safety systems executive, Johannes Heidecke, departed after a company reorg. (Source)
• Apple Is Suing OpenAI for Allegedly Stealing Hardware Secrets — Apple launched a lawsuit against OpenAI and a hardware exec for allegedly misusing confidential designs. (Source)
• OpenAI GPT-5.6: AI Could Do Anything, Then It Met ARC-AGI-3 — A close look at where GPT-5.6 shines and where it falls short on a major general intelligence benchmark. (Source)
• Quoting OpenAI — Details on how cloud and desktop “Work” conversations in ChatGPT are managed and stored. (Source)
• The new GPT-5.6 family: Luna, Terra, Sol — Context and industry reaction to the GPT-5.6 launch and three-model approach. (Source)
• Introducing Muse Spark 1.1 — Meta releases new Muse Spark 1.1 model with its first public API and improvements for agentic tools. (Source)
• llm-meta-ai 0.1 — A new tool lets users run prompts against Meta’s new AI models. (Source)
• llm 0.31.1 — Bug fix release for the “llm” command-line AI prompt tool. (Source)
• Rewriting Bun in Rust — The Bun JavaScript runtime is moving from Zig to Rust for performance/maintainability. (Source)
• Quoting Kenton Varda — Some teams are now banning AI-written change descriptions due to quality issues. (Source)
• sqlite-utils 4.0, now with database schema migrations — sqlite-utils 4.0 adds major schema migration features, making the open-source DB tool more powerful. (Source)
• tencent/Hy3 — China’s Tencent releases “Hy3,” a giant new open-source Mixture-of-Experts AI model. (Source)
• Secret Claude tracker shocks users after Anthropic’s anti-surveillance stance — Anthropic removed hidden tracker code from Claude Code after a privacy backlash in China. (Source)
• Anthropic found a hidden space where Claude puzzles over concepts — New research reveals how Anthropic’s Claude model organizes concepts internally. (Source)
• Your family’s $300 stake in OpenAI — Analysis of Sam Altman’s proposal for a universal AI dividend for Americans. (Source)
• OpenAI’s CEO of AGI Deployment, Fidji Simo, Is Stepping Down — AGI deployment chief at OpenAI leaves full-time role, stays as adviser. (Source)
• Anthropic Wants You to Pay Up for Claude Fable 5 — Claude’s newest model will only be available to paid Anthropic customers. (Source)