GPT-5.6 Explained: Sol vs Terra vs Luna, ChatGPT Work and the New Desktop App

ChatGPT used to be the simple one. You opened it, you typed, it answered. That era ended on 9 July 2026, when OpenAI shipped three new models with astronomical names, a new agent called ChatGPT Work, a new desktop app that swallowed Codex whole, and a settings menu where you now pick a model, an effort level and a speed before you have typed a word.

Two weeks ago I wrote a plain-English guide to Microsoft's Copilot Cowork pricing because nobody could follow it. I did not expect to write the same article about ChatGPT quite so soon. But here we are, and the questions filling my inbox are the same ones: what actually changed, which model should we use, what does the $20 plan really include, and should we switch, stay, or wait?

So I have done the same boring work again. I have read the release notes and the rate cards, run both tools side by side, and turned it into the guide I wish OpenAI had published. Here is the plain-English version for anyone who has to make a decision about it.

What actually launched on 9 July

Three things arrived at once, and most of the confusion comes from treating them as one thing.

• Three new models. GPT-5.6 comes as Sol, Terra and Luna, each with a different capability level and a very different running cost.
• ChatGPT Work. A new agent mode that pursues finished outcomes, such as research reports, spreadsheets, presentations and even websites, rather than chat replies.
• A unified desktop app. The old Codex app became the new ChatGPT desktop app, containing Chat, Work and Codex as three modes. The previous desktop app lives on as "ChatGPT Classic".

The cleanest way to hold it in your head is that ChatGPT is no longer one product. It is three working modes in one platform, and each mode has a different relationship with your files, your apps and your usage allowance.

ChatGPT Work is the part that matters for most teams. It is not a stronger chat window. It can pursue a multi-step outcome, pull in files and connected apps, ask questions midway through the job, accept a change of direction, request approval before consequential actions, and hand back an editable deliverable at the end. That is the same territory Claude Cowork has occupied since early 2026, which is exactly why the comparison later in this article matters.

GPT-5.6 Sol vs Terra vs Luna: which model should you use?

OpenAI's stated logic for the names is that the generation number stays, while Sol, Terra and Luna are durable capability tiers that can advance on their own cadence. In practice: the Sun, the Earth and the Moon, in descending order of firepower and cost.

The credit ratios are exact, and they are the single most useful fact in this whole launch. Per million output tokens, Sol costs 750 credits, Terra 375 and Luna 150. Output is also six times more expensive than fresh input on every model, which means long, polished answers and multiple alternative versions quietly cost more than the reading the model did to produce them.

Sol is for when being wrong is expensive. Ambiguous questions, a dozen conflicting sources, work going in front of a client or a board, code where the bug is subtle. If verifying the answer would itself take you hours, pay for Sol.

Terra is the practical default, and I suspect the commercially important one. First drafts, spreadsheet cleaning, meeting synthesis, routine coding, document transformation, connected-app workflows with a clear objective. Strong reasoning without Sol's bill.

Luna is the operational model, not the cut-down one. When the output is already defined and the value is speed, consistency and volume, Luna is the right answer, not the cheap answer. Extract five fields from 200 documents. Categorise the tickets. Convert the notes into the template. Every one of those on Sol is money burned for no extra quality.

One detail that catches people out: Terra and Luna are not selectable in ordinary chat. Standard ChatGPT conversations still run on GPT-5.5 Instant for quick replies, with Sol powering the reasoning options. The full three-model choice only exists in Work, Codex and the API. If you have been hunting for Luna in the everyday model picker, you can stop.

Why is ChatGPT naming models after the solar system?

Because the old names were worse. Nobody, including people who train teams on these tools for a living, could reliably remember whether o3 was better than 4o, or where "mini-high" sat in the hierarchy. Named capability tiers that persist across generations are genuinely better product design, and I will happily say so.

That said, the industry had fun with it. Within hours of the announcement, Solana's official account replied with "Sam Altcoinman", because Sol is a crypto ticker and Terra and Luna are forever associated with one of the most spectacular collapses in crypto history. OpenAI has, presumably by accident, named its two cheaper models after roughly forty billion dollars of vaporised cryptocurrency. The models are excellent. The naming committee may want a wider whiteboard next time.

ChatGPT effort settings explained: Power, model, effort and speed

Here is the part that makes people's eyes glaze over. In Work and Codex, you are no longer choosing one thing. You are choosing up to three, and there is a simplified preset on top called Power. This is what each control actually does.

The effort ladder runs Light, Low, Medium, High and Extra High, and then two special modes: Max, which gives one model even more thinking time for a single very hard task, and Ultra, which coordinates four agents in parallel and behaves less like a harder-thinking model and more like a small project team dividing the work. Higher effort helps when a task genuinely needs planning and trade-off analysis. It adds almost nothing to simple extraction or reformatting except time and burned allowance.

If you multiply it out, as the researcher Sebastian Raschka did in launch week, the models, efforts and speeds combine into around 72 possible configurations. Nobody needs 72 configurations. You need three habits:

• Start with Terra on Medium. This should be your team's stated default, not Sol, which is what the out-of-the-box Power preset gives you.
• Escalate to Sol, or raise effort, when the task's ambiguity, value or risk justifies it. One deliberate upgrade, not a reflex.
• Drop to Luna on Light for anything repetitive and well-defined. This is where allowances quietly stretch to five times their apparent size.

What the $20 Plus plan actually includes

This is the section to forward to whoever approves the invoices, because three different limits are being confused in almost every discussion I have seen.

• Ordinary chat keeps its familiar message limits. Nothing dramatic changed there.
• Work and Codex share a separate agentic allowance, measured in credits behind the scenes, refreshing on a five-hour window, with weekly caps also possible.
• The 5x and 20x multipliers describe the Pro tiers, not Plus. The $100 Pro plan is five times the Plus allowance, and the $200 Pro plan is twenty times. There is no 20x anything on the $20 plan.

The key sentence, and the one OpenAI's own marketing will not put this bluntly: Plus does not buy a fixed number of tasks or tokens. It buys a variable allowance whose real size depends on which model you choose and how much work each task does. A task's consumption moves with the model, the effort, the context it reads, the tools it calls and the length of what it writes. OpenAI publishes planning ranges rather than promises, and the ranges are wide.

When you hit the limit, you have four options: wait for the window to reset, switch to a smaller model or lower effort, buy additional credits where the feature supports it, or, for technical local work, plug in an API key and pay standard API rates. Included usage is consumed before purchased credits, and credit availability varies by feature, so check the usage dashboard rather than assuming everything continues as pay-as-you-go.

These are US list prices throughout, so expect UK billing to track the dollar rate plus VAT, the same as it does today.

The new ChatGPT desktop app, and what happened to Codex

The short version: the Codex desktop app became the new unified ChatGPT app, with Chat, Work and Codex inside it. Codex did not disappear; it is a mode. Existing Codex projects and settings carry over, and long-time Codex users can even keep it as their default opening view. The previous ChatGPT desktop app stays installed as ChatGPT Classic and keeps receiving model updates, though the new agent features belong to the new app.

Why the desktop app matters is what it can reach. The web version of Work runs in the cloud and works with uploads and connected apps. The desktop version can, with permission, work with your local files and folders, control supported applications, and drive Excel directly through an add-in. There is one launch limitation worth knowing before you rely on it:

Beyond the headline features, the launch quietly shipped a lot: ChatGPT Sites turns a Work task into an interactive website or lightweight app, though OpenAI states it is not available in the EEA, Switzerland or the UK at launch, so UK readers should not budget around it yet. Scheduled and monitoring tasks can run jobs on a timetable or watch for changes, capped at ten active tasks. Plugins replace the old App Directory, packaging apps, skills and templates into installable workflows. And a cloud browser lets Work navigate public websites when no connected app can do the job, though it cannot log in to anything, use passwords or complete payments, which is the right call for something this new.

The token limit question: ChatGPT vs Claude for organisations

For organisations this is the comparison that actually matters, and it is also where two different limits get muddled constantly, so let me separate them before the numbers. The context window is how much a single conversation can hold: the hundred-page contract, the full board pack, the codebase read in one pass. Usage is how many messages or tasks you can run before a cap kicks in. They are different limits, the two tools win one each, and mixing them up is how the wrong product gets bought.

On context, I went into this research expecting ChatGPT to win comfortably. The official documentation says otherwise, and it is worth being precise, because I have now seen this assumed incorrectly in three separate procurement conversations.

On the matched $20 plans, the gap is wide and it runs in Claude's favour. ChatGPT Plus gives 54K tokens of context for everyday responses and 256K when a reasoning model is used. Claude Pro gives 500K on most current models and a full 1M tokens with Sonnet 5, in ordinary chat. Even ChatGPT's $200 Pro plan, at 128K and 400K, offers less in-chat context than Claude's $20 plan. GPT-5.6 does support a 1M window, but only through the API, where developers pay per token and where Claude offers the same.

Before anyone concludes the whole question is settled, the honest picture has a second half. Where ChatGPT genuinely leads is usage volume and transparency. Plus advertises effectively unlimited everyday chat messages, while Claude enforces five-hour session windows plus weekly caps that heavy users do hit, and Anthropic publishes multipliers rather than absolute numbers. OpenAI publishes actual planning ranges and a credit rate card for its agentic work; with Claude you largely discover your real ceiling by using it. So the accurate one-line summary for a budget holder is this: Claude gives you far more context per conversation at the same price; ChatGPT gives you clearer, better-documented usage economics. Which of those matters more depends on whether your bottleneck is document size or task count.

ChatGPT Work vs Claude Cowork: the full comparison

I use Claude Cowork daily in my own business and have spent the days since launch running ChatGPT Work alongside it, so this comparison comes from both sides of the fence, not from two press releases.

Connectors: count the right thing

On raw directory size, ChatGPT is ahead: community counts put its app directory in the several hundreds of verified integrations, against roughly three to four hundred in Claude's connectors directory. But raw counts are the wrong thing to count. Claude additionally supports desktop extensions and any custom MCP server, which is how technical teams wire in their own internal systems, and both tools reach thousands more apps through Zapier.

The Outlook story is the perfect illustration of why the details matter. ChatGPT ships Outlook Email and Calendar connectors that work with ordinary Microsoft accounts, and having your inbox available inside an AI workspace is one of those features you cannot un-want after a week of using it. Claude does now have a Microsoft 365 connector covering Outlook, Teams, SharePoint and OneDrive, with write access, and on paper it is the deeper integration. The catch: it requires a Microsoft business tenant, so anyone on a personal Outlook account, which includes a lot of freelancers and small businesses, is excluded. If you have been wondering why Outlook works in one tool and not in yours, that is the reason.

Speed and reliability: the honest state of play

Claude's desktop app has a documented performance problem, and pretending otherwise would help nobody. Public bug reports describe multi-gigabyte virtual machine bundles that never clean themselves up, noticeable idle CPU usage and slow startup, and my own experience matches: Cowork sessions can feel heavy, and I have learned to plan around it. Early users have been vocal that the new ChatGPT app feels snappier. The fair caveat is that ChatGPT Work is days old. Claude Cowork has more than a million sessions across hundreds of thousands of organisations behind it, and there is no independent head-to-head benchmark of the two agents yet. Fast and unproven versus slower and battle-tested is a real trade-off, not a rhetorical one.

So should your organisation switch?

This is the question underneath every AI tooling conversation I have had this year. One vendor ships, the other leapfrogs, and every release makes someone's six-week-old rollout decision look premature. So let me say the most important thing first: a launch is not a reason to migrate. Switching costs are real: retraining, rebuilt workflows, new governance, lost muscle memory. The moment to switch is when the other tool is decisively better for your specific task mix over a full quarter, not for the fortnight after a keynote.

With that said, here is how I would call it by situation:

• Already on ChatGPT and happy: stay, and adopt Work deliberately. Set Terra as the default, write a one-page model policy, and pilot with a small group before rolling wide.
• Already on Claude and happy: stay. Nothing in this launch removes Claude's advantages in context size, local desktop work and extensibility. Watch how Work matures, and reassess in a quarter with your own usage data.
• Google Workspace organisation: ChatGPT Work has the edge, with native Docs, Sheets and Slides output.
• Microsoft-heavy organisation: closer than the headlines suggest. Claude's M365 connector is deeper if you have a business tenant; ChatGPT's Outlook and Excel story works for ordinary accounts. And if you are weighing Copilot itself into this decision, I compared its pricing model in the previous article.
• Individuals and small teams choosing fresh: run one identical real task through both $20 plans for a fortnight and count what you actually consumed. Your task mix will answer the question better than any comparison table, including mine.

The caveat before you paste company data into anything

A $20 personal subscription is not a governed business workspace, on either side. On personal ChatGPT plans your content can be used to improve OpenAI's models unless you switch that setting off, and third-party plugins carry their own permissions and policies. The Business and Enterprise tiers exclude workspace content from training by default and add proper admin controls; Anthropic draws a comparable line between its consumer and commercial products. If the work involves confidential information, the plan tier and the settings matter more than the logo on the app.

What I would do this month

1. Do not let the Power preset choose for you. The default is Sol on Medium, which is the most expensive routine setting. Make Terra your team default and write it down.
2. Match the model to the stakes, not the deadline. Sol for high-stakes and ambiguous work, Terra for the everyday, Luna for volume. The ratios are 1x, half, one fifth. That is the entire policy.
3. Treat the first month as research. Watch the usage dashboard, note what heavy tasks consume, and only then decide whether anyone needs a Pro tier.
4. Keep approval steps on consequential actions. Agents that can browse, fill forms and control applications are powerful and new. Keep a human on the send button.
5. Reassess the landscape quarterly, not per keynote. You will otherwise spend more on switching than you ever spent on tokens.

The tools have stopped competing on whether they can do the work. They now compete on how the work is priced, governed and fitted into your stack. That is a procurement skill, not a prompt skill, and it is the one most teams have not built yet.

If your team is adopting ChatGPT, Work or any of this month's alphabet soup and you want it to land as capability rather than confusion, that is exactly what I do: hands-on ChatGPT training for teams, built around your real tasks rather than vendor demos.