🔥 This Week’s Highlights
This was the week the assistant stopped answering and started working. OpenAI put GPT-5.6 in everyone’s hands and shipped ChatGPT Work — an agent you hand a goal, not a question, and it comes back hours later with a finished deck. Anthropic pushed Cowork onto your phone so those long-running jobs follow you out of the house, and xAI and Meta both dropped agent-first models within 24 hours of each other. Meanwhile OpenAI’s voice mode learned to say “mhmm,” and Apple sued OpenAI over its unreleased hardware. Busy week.
🤖 AI Models & Features
ChatGPT Work: OpenAI’s Agent That Delivers Finished Docs, Slides and Sheets
🧑💼 Products & Platforms · High Priority · 30 min to try
This is the one. On July 9, OpenAI shipped an agent that takes an outcome rather than a prompt — “build me a competitor comparison deck” — then gathers context from your connected apps, breaks the job into steps, and grinds on it for hours. What comes back isn’t a chat response; it’s a spreadsheet, a slide deck, a report, a working web app. Rolling out on Pro, Enterprise and Edu first, then Plus and Business. Try it on something you’d otherwise burn an afternoon on, and be honest with yourself about the result.
GPT-5.6 Ships to Everyone — and Codex Folds Into a New ChatGPT Desktop App
🚀 News & Announcements · High Priority · 5 min to try
GPT-5.6 finished its public rollout on July 9 in three tiers: Sol (flagship), Terra (matches GPT-5.5 at roughly half the price) and Luna (fastest, cheapest). It’s already your default in ChatGPT — you don’t have to do anything. The quieter but more useful news: the standalone Codex app is merging into a new ChatGPT desktop app that puts Chat, Work and Codex in one window, on every plan, including Free. Go download it.
GPT-Live: ChatGPT’s Voice Mode Can Now Listen and Speak at the Same Time
🎙️ Products & Platforms · High Priority · 5 min to try
Old voice mode was a walkie-talkie: you talk, it talks, repeat. GPT-Live (July 8) is full-duplex — it processes what you’re saying while it’s speaking, making a decision many times per second about whether to talk, listen, pause or interrupt. It backchannels with “mhmm.” It lets silence sit while you think. And for hard questions it quietly hands off to a frontier model in the background and returns with the answer mid-conversation. It’s rolling out globally and it genuinely changes what voice mode is for.
Claude Cowork Comes to Web, iOS and Android
🤝 Products & Platforms · High Priority · 15 min to try
Anthropic’s agentic workspace escaped the desktop on July 7. Cowork sessions now run in the cloud and follow you across devices — start a task at your desk, close the laptop, check progress from your phone on the train. Connectors, skills, scheduled tasks and file previews all work on mobile. The desktop app keeps the superpowers (local files, browser control, computer use). Buried in the announcement is the best stat of the week: across 1.2 million Cowork sessions, the overwhelming majority had nothing to do with writing software. Usage limits are extended through August 5.
Grok 4.5: xAI’s ‘Opus-Class’ Model, Trained on Real Cursor Sessions
⚡ News & Announcements · Medium Priority · 15 min to try
xAI’s first release since going public and swallowing Cursor, out July 8. Built on a 1.5-trillion-parameter foundation and — this is the interesting part — trained on real Cursor session data, meaning actual humans doing actual coding work. Musk calls it “Opus-class but faster and cheaper,” and at $2/$6 per million tokens versus Opus 4.8’s $5/$25, the price claim is at least true. Available in Grok Build and in Cursor on all plans. Not in the EU yet.
Meta Muse Spark 1.1 — and Meta’s First Paid AI Model
🧩 Products & Platforms · Medium Priority · 30 min to try
Meta released an agentic multimodal model on July 9 with a 1M-token context window that can use computers, coordinate sub-agents, and hold a long project together. Capable, but that’s not the headline. The headline is that Meta — the company that built its AI reputation on giving Llama away — put a price on it: $1.25/$4.25 per million tokens via the new Meta Model API, with $20 in free credits. Under Alexandr Wang, the open-weights evangelist has started charging rent.
Claude Reflect: A Monthly Recap That Nudges You to Use AI Less
🪞 Products & Platforms · Medium Priority · 5 min to try
Anthropic shipped a “Claude Wrapped” that occasionally suggests you close the tab. Reflect (Settings → Reflect, beta on Free/Pro/Max, requires memory on) shows your top topics, most active day, peak hour, and scores your usage against a four-part AI fluency framework — Delegation, Description, Discernment, Diligence. It also lets you set quiet hours and break reminders. Slightly uncomfortable in the way a step counter is uncomfortable. Worth five minutes.
GPT-5.6 Sol Reportedly Cracks a 50-Year-Old Math Conjecture in Under an Hour
🧮 Concepts & Research · Medium Priority · 5 min to try
Reported July 11: GPT-5.6 Sol Ultra produced a proof of the Cycle Double Cover Conjecture, open for half a century, in under an hour — using 64 subagents working in parallel. Hold the result loosely until mathematicians verify it. But hold onto the method, because that’s not a hint, it’s a preview: parallel subagents swarming one hard problem is the same architecture that makes ChatGPT Work able to “think for hours.”
GPT-5.6 Sol, Terra and Luna Land in GitHub Copilot
🐙 Coding Tools · Medium Priority · 15 min to try
Same-day availability in Copilot, which is faster than these integrations usually move. Even if you’re not a developer — if you touch scripts, formulas, or automations — the model picker in Copilot Chat now matters: Sol for genuinely hard problems, Terra as the sensible default, Luna when you just want it fast. One dropdown, real difference.
NYT/Cambridge: Extremist Groups Are Using Every Major AI Chatbot
⚠️ Concepts & Research · Medium Priority · 5 min to try
A Cambridge study, reported by the NYT this week, documents Boko Haram members using ChatGPT, Claude and Gemini to plan attacks, design explosives and maintain weapons — with ISIS operatives training commanders on the tools. This is the clearest public evidence yet that consumer safety filters are being routinely worked around. Expect the downstream effect to land on you: tighter refusals, more friction on legitimate chemistry and security questions, and fresh pressure on open-weight releases where filters can simply be deleted.
Mistral Leanstral 1.5: A Free, Apache-2.0 Model That Proves Code Correct
🔬 Concepts & Research · Low Priority · 1 hour+ to try
The week’s notable open-weights drop. Leanstral 1.5 is a 119B MoE (6B active), fully Apache 2.0, that writes Lean 4 proofs — it doesn’t just generate code, it mathematically proves the code does what it claims. It solved 587 of 672 PutnamBench problems and found five previously unknown bugs across 57 real repositories. Weights on Hugging Face, free API. Niche today; this is where “AI-written software you can actually trust” has to end up.
Indeed Data: Software Job Postings Are Up 15% Since Claude Code Launched
📈 Concepts & Research · Low Priority · 5 min to try
Indeed’s Hiring Lab reported on July 11 that US software development postings have grown ~15% since Claude Code launched in February 2025 — while overall job postings fell 7%. Correlation isn’t causation and postings aren’t hires, so don’t overclaim. But if “AI is killing developer jobs” has been a settled premise in your head, this is the most credible thing yet arguing the opposite.
Every New Car in the EU Now Has AI Watching the Driver
🚗 News & Announcements · Low Priority · 5 min to try
Since July 7, every newly registered car in the EU must include a driver distraction detection system — a camera reading your gaze and head movements to catch inattention, processed on-device and never sent to auth…