The 60-second briefing
The five stories everyone will be talking about this week — biggest first:
GPT-5.6 is now generally available. OpenAI launched Sol, Terra, and Luna across ChatGPT, Codex, and the API, plus an ultra setting that coordinates agents for hard tasks. The real story is price-performance: smaller models are meant to make useful intelligence cheaper.
ChatGPT Work is turning chat into a place where work finishes. The agent can pull context from connected apps, make docs, sheets, slides, and web apps, and keep projects moving with scheduled tasks. The shift is from answering to doing.
GPT-Live makes voice feel less like taking turns. OpenAI’s new full-duplex voice models can listen and speak at the same time, pause, interrupt, and delegate deeper work in the background. That sounds small until you try talking to an AI without waiting for it to finish every sentence.
Claude is giving users a mirror. Anthropic’s new Reflect beta shows how you use Claude across the last 1, 3, 6, or 12 months, including topics and recurring task patterns. It is an unusually honest product move: measuring whether AI use is helping instead of only encouraging more of it.
Claude Science is packaging research into one workbench. The app connects literature search, scientific tools, computing resources, and auditable artifacts. The goal is fewer handoffs between databases, notebooks, terminals, and figures.
The best small story this week came from a maker who was tired of being chained to a desk while an AI coding agent worked.
So Salvatore Castellitti built CodeMote. It pairs an iPhone with the computer or VPS where Claude Code, Codex, or another terminal agent is running.
You see the live terminal on your lock screen. When the agent needs approval, your phone tells you. The agent keeps running on the machine, so your connection can disappear without killing the job.
It is one specific annoyance removed by the person who had it. That is my favorite category of AI progress.
What everyday people just built
Typeahead made autocomplete local and app-aware. Typeahead 2.0 suggests text in any Mac app, works offline, supports writing styles per app, and keeps the model on-device. It is a $79 one-time purchase. See the launch.
AI Emaily treated the inbox like a chief-of-staff problem. It triages messages, drafts replies in your voice, and supports Manual, Copilot, and Autopilot modes with approval before sending. The useful bit is the control dial. See the launch.
Katalyst put an agent on top of Salesforce. It turns calls, emails, and calendars into records, follow-ups, account plans, and meeting briefs. Reps spend time selling, not reconstructing deal history for the CRM. See the launch.
AnySearch built search for agents instead of people. It searches trusted sources in parallel, filters duplicates and SEO spam, and returns structured context through a skill, MCP, or API. Better input makes agents less confidently wrong. See the launch.
CodeMote turned a long-running coding agent into a mobile workflow. It adds live terminal access, push notifications, diffs, and approvals from an iPhone. The agent stays on your machine; the phone is the window. See the launch.
Notice the pattern: the newest useful AI apps are attaching themselves to surfaces people already use — a Mac text field, an inbox, Salesforce, search, or the phone in your pocket.
The money: who just raised
The infrastructure bill is getting bigger because companies want control over how their agents run.
Prime Intellect — $130M Series A, led by Radical Ventures. It is building compute, reinforcement learning, and evaluation tools so companies can train their own agents. Read more.
SambaNova — $1B Series F first close, led by General Atlantic at an $11B valuation. The chip company is targeting fast, private inference for enterprises and governments. Read more.
Ollama — $65M Series B, led by Theory Ventures. The open-source tool helps developers run open-weight models locally and says it reaches nearly 9 million developers monthly. Read more.
Gradium — $100M seed, with Nvidia among the new investors. The Paris startup is building ultra-low-latency voice models. A seed round. $100 million. Read that twice. Read more.
Try this before Sunday
Use ChatGPT Work for one task you normally scatter across five places: a budget review, launch plan, messy research folder, or sales recap.
Give it the source files, define “done,” and ask it to show its work before editing or sending. Stop spending judgment on file hunting and formatting.
Review these files and produce one decision memo. First, list the important facts with links back to the source. Then identify the three decisions I need to make, the risks or missing information, and a recommended next step for each. Do not invent facts. Keep the final memo under 600 words and ask me before changing or sending anything.
The better question is whether it can carry real work from scattered inputs to a useful next decision.
— John
P.S. I read every reply — the best idea gets featured next week.