The AI Catchup -- July 15, 2026

Welcome back. Last issue was about models: GPT-5.6 went GA with prices and a benchmark table. This issue is about what those models are being wrapped in. OpenAI quietly restructured ChatGPT itself around a third agent called Work, Cursor wired the whole GPT-5.6 family into its model picker and shipped the most practical context-management feature we have seen in months, and Anthropic launched a dashboard that tells you how you actually use Claude. Everything this week is a rollout you can act on, so we are leading with a table that answers the only question that matters: can you use it today?

Let us get into it.

At a Glance

This Week in AI

ChatGPT is no longer one thing. OpenAI introduced ChatGPT Work (vendor post July 9), and with it a three-way split: Chat for questions and conversation, Work for longer research and finished materials, and Codex for software development. Work is powered by Codex and GPT-5.6, and per OpenAI it can take action across your apps and files and stay with a project for hours to turn a goal into a finished deliverable -- a document, spreadsheet, presentation, report, or Site.

The decision rule is simple: choose Work when the output is a deliverable, choose Codex when the output is code. Work follows the same usage structure as Codex, and through Scheduled Tasks it can run once, repeat on a schedule or trigger, or monitor for changes -- which makes it a standing job runner, not a one-shot generator.

Two operational details worth knowing before you lean on it:

• The desktop app is the generous surface. Chat, Work, and Codex are on every desktop plan including Free, globally. On web and mobile, Work went to Pro, Enterprise, and Edu first, with Plus and Business following over days.
• File access depends on where Work runs. Web and mobile Work runs in the cloud and cannot reach files on your computer. Desktop Work can use local files and apps with your permission, and at launch, cloud Work conversations do not appear in desktop Work.

The bigger pattern: two weeks ago Codex landed in the ChatGPT mobile app, last week GPT-5.6 went GA, and now the consumer product is reorganized around agents as the top-level navigation. The chat box is becoming the smallest thing ChatGPT does.

Cursor Had the Most Usable Week

Two Cursor updates pair well enough that we are treating them as one story. First, the GPT-5.6 family arrived in Cursor (vendor post July 9): Sol for difficult coding problems and complex agent workflows, Terra for everyday development, and Luna for fast, cost-efficient edits, all selectable from the model picker. Cursor published one benchmark datapoint -- Sol scores 67.2% at Max effort on CursorBench 3.2 -- so run your own comparison on your own tasks rather than taking a single number on faith.

Second, and the one we think you will use daily: Side chats (vendor post July 10). Type /side or /btw (or hit the plus button at the top of the chat panel) and Cursor opens a separate agent conversation that inherits context from your main chat. The key design choice is that a side chat is not a throwaway scratchpad -- it is a durable, full agent conversation you can revisit later and @-mention to pull its context back into the main thread. Ask "what are the alternatives here?" without pushing your main agent off track, then feed the answer back in when it matters.

Ship It This Week

Three direct wins you can act on today.

Branch instead of derailing. Next time you are mid-agent-run in Cursor and need to check a tangent -- what a file does, whether an approach is sane, what the alternatives are -- type /btw and ask there. The side chat inherits your context, your main run stays on track, and if the tangent turns into a real design discussion, @-mention it back into the main thread. This is a zero-cost habit change that fixes the most common way long agent sessions go sideways.

Run a Sol bake-off on one real task. Pick your longest-running recurring task -- a multi-file refactor, tool-heavy debugging -- and run it once on GPT-5.6 Sol from Cursor's model picker and once on your current default, same prompt. Last issue gave you Sol's Terminal-Bench numbers and API prices; this is the week you find out what it does on your codebase.

Hand a recurring chore to ChatGPT Work. Work is on every desktop plan, including Free. Pick one recurring research deliverable -- a weekly competitor summary, a market scan, a formatted report -- and set it up with Scheduled Tasks so it repeats on a schedule. Two cautions from the launch notes: use desktop Work if the task needs local files, and approve important actions as it runs rather than letting it free-run.

Quick Hits

• Anthropic launched Reflect, a usage recap for Claude. The beta dashboard (vendor post July 8) lives in Settings and shows when you use Claude most and what you worked on, over the past 1, 3, 6, or 12 months, with optional quiet hours and break nudges. It is on Free, Pro, and Max (web and desktop) and requires memory on. The privacy exclusions are the load-bearing detail: no incognito chats, no files from connected tools, and health integration conversations are left out entirely. If you have ever wondered where your Claude hours actually go, this answers it in one screen.

---

That is it for this issue. The through-line this week is packaging: the models settled last issue, and now the vendors are fighting over the surfaces you use them through -- three agents in ChatGPT, a fuller model picker and branching threads in Cursor, and a usage mirror in Claude. If a teammate still thinks ChatGPT is just a chat box, forward this -- the Work split changes what it is for.

Until next week, stay caught up.