attentionfactory
All guides

Attention Factory · Field Guide · July 2026

The ChatGPT upgrade you can actually use

GPT-5.6 shipped as three models. ChatGPT Work shipped as an agent that finishes jobs. Sites shipped as a way to publish what it builds. Here is what each one is, who it is for, and the exact prompts to run today.

Start here

The one-paragraph version

OpenAI stopped shipping one model and started shipping a lineup. GPT-5.6 comes in three tiers, so the skill now is picking the right one instead of defaulting to the biggest. Alongside it, ChatGPT Work turns the chat window into an agent that plans a project, touches your connected apps, and hands back finished files. Sites lets it publish a working web app or dashboard at a shareable link. The upgrade is not that ChatGPT answers better, it is that ChatGPT now delivers.

Part one

GPT-5.6: Sol, Terra, Luna

The number is the generation. The names are capability tiers that can improve on their own schedule, which means "Sol" will keep meaning "the flagship" even after the version number moves. Stop thinking in versions and start thinking in lanes.

Sol Flagship · deepest reasoning

The one to reach for when the task needs persistence: long-running agent work, big codebases, research that spans dozens of steps, anything where you would rather wait and get it right. In ChatGPT it powers the Medium, High, and Extra High reasoning settings on paid plans, with a Sol Pro option on Pro and Enterprise.

Terra Balanced · everyday default

Roughly the performance you were getting from the previous flagship, at a fraction of the price. This is the sane default for scoped work: drafting, first-pass review, everyday agentic tasks, most content and ops workflows. Free and Go users get Terra inside Work and Codex.

Luna Fast · lowest cost

The volume lane. Summarising, labelling, extracting, classifying, cleaning data, first drafts you will rewrite anyway. If the job is mostly shuffling text rather than thinking hard about it, Luna does it fast and cheap, and you escalate only when it stalls.

How to route your own work

Your taskLaneWhy
Long agent run, multi-file code, deep researchSolHolds the thread across many steps without losing the plot
Content drafts, scoped features, reviews, analysisTerraFlagship-class output at everyday cost
Bulk summaries, tagging, extraction, cleanupLunaSpeed and price are the constraints, not depth

The design jump

The visible upgrade for most people is what comes out the other end. GPT-5.6 produces noticeably better presentations, documents, and spreadsheets, including fully editable decks built from a prompt and your own source material, with real layout and hierarchy instead of a wall of bullets. It also follows your templates and reference files far more closely, which is the difference between a deck you rebuild and a deck you tweak.

Prompt · editable deck from source material
Build a 10-slide investor update from the attached notes and last quarter's deck.
Match the attached deck exactly: same fonts, same colour palette, same slide
structure, same tone of voice.
One idea per slide. Every number must trace back to a line in my notes.
Where a number is missing, leave a visible placeholder instead of guessing.
Give me the editable file, not an image.

Part two

ChatGPT Work: the agent that finishes the job

Until now ChatGPT had two doors: Chat for conversation, Codex for code. Work is the third door, and it is aimed at everyone who is not a developer but wants the muscle of a coding agent.

You give it an outcome. It gathers context from your connected apps and files, breaks the outcome into steps, and works through them, staying with a project for hours if that is what the job takes. It runs on a virtual machine in the cloud, so it keeps working when your laptop is shut. It hands back finished artefacts: docs, spreadsheets, slides, reports, and web apps.

What makes it different from asking ChatGPT a question

  • It holds context across the whole chain. You do not re-explain the project at every stage.
  • It acts, not just answers. Connected apps mean it can read your email, calendar, Slack, notes, and repos and then produce something from all of them at once.
  • It asks before it commits. You set what it can touch, when it checks in, and which actions need your sign-off. Plan mode shows you the plan before any work starts.
  • It moves between devices. Kick a task off on your phone, review the draft in a queue, pick it back up on desktop.
  • Scheduled Tasks make it recurring. It can run once, run on a schedule, run on an event, or watch something and tell you when it changes.

Your first three tasks

Pick work you already do weekly. The point of the first run is calibration, not magic.

Task 1 · the inbox triage
Go through my email, calendar and Slack from the last 7 days.
Pull out anything that needs a decision or a reply from me, and ignore anything
that is FYI only.
Build me a priority list with three buckets: reply today, reply this week, no
action needed but worth knowing.
For each item in bucket one, draft the reply in my voice and leave it as a draft.
Do not send anything. Check in with me before you touch any draft.
Task 2 · the recurring report
Every Friday at 4pm, pull this week's numbers from the connected sheet and the
analytics dashboard.
Compare against the previous week and the same week last month.
Write a one-page summary: what moved, why it likely moved, and the one thing I
should act on.
Chart the two metrics that moved the most.
Deliver it as a doc and drop a three-line version in my Slack DMs.
Task 3 · the messy project
Here is a folder of raw material for a campaign: notes, transcripts, and two
competitor pages.
Read all of it first, then show me your plan before you build anything.
The output I want is: a one-page campaign brief, a content calendar as a
spreadsheet, and a 6-slide deck for the client.
Use my brand template for the deck. Ask me questions where the source material
is thin instead of inventing an answer.

Guardrails worth setting on day one

  • Turn Plan mode on and read the plan before approving. This is where you catch a misread brief, and it costs you thirty seconds.
  • Keep approvals on for anything that leaves your account: sending mail, posting, publishing, paying.
  • Connect apps one at a time and watch what it pulls in. Full access on day one gives you no way to trace a bad output back to a source.
  • Write down how long the task takes you by hand before the agent touches it. Without that number you cannot tell whether this is saving you anything.
  • Give it a task you can grade. If you cannot instantly spot a wrong answer, you are not supervising, you are hoping.

Part three

Sites: publish what it builds

Sites turns a description into an interactive site or web app at a shareable URL, with the code, the backend pieces, and the deployment handled for you. It is in public beta, and it is the fastest path from "I have information" to "my team can use it."

What it is actually good for

  • Live dashboards that stay updated as the underlying data changes.
  • Project trackers and launch calendars your team opens instead of asking you for status.
  • Prototypes you can put in front of a client on a call rather than describing.
  • Internal portals onboarding hubs, resource libraries, SOP pages.
  • Interactive reports where the reader can filter and explore rather than scroll a PDF.
Prompt · internal tracker in one shot
Build me a web app my team can open at a link.
It is a content tracker: each item has a title, platform, owner, status
(idea / scripting / filming / editing / scheduled / posted), and a publish date.
Give me a kanban view by status and a calendar view by publish date.
Anyone with the link can add and edit items. No login.
Seed it with the 12 items in the attached sheet.
Keep the design clean: dark background, one accent colour, big readable type.

The move that separates the people who get value here from the people who post a screenshot: build the thing you would otherwise have chased people for. Every recurring "where are we on this?" message is a Site waiting to be made.

Part four

The 30-minute setup

  1. Update your app. Desktop is where the agent is strongest: it can reach local files, use a built-in browser, and work across apps. Mobile is for kicking off and reviewing.
  2. Connect one app. Start with the one holding the context you repeat most often. For most people that is email or Drive.
  3. Pick your default lane. Terra for everyday. Sol when the task is long or the stakes are real. Luna for volume.
  4. Run one known task. Use Task 1 above. Grade the output honestly against what you would have produced.
  5. Schedule one recurring job. The weekly report, the Monday agenda, the daily dashboard check. Recurring work is where the hours actually come back.
  6. Publish one Site. Turn your most-asked-about status update into a link.