GPT-5.6 isn’t the real news — the new ChatGPT app is

![[featured.png]]

Everyone spent launch week arguing about benchmarks. Sol beats Fable here, Fable beats Sol there. I read the charts too — and then I started looking at the application itself.

The app suggests a new way of routing work: Chat for an immediate answer, Work for a finished deliverable, and Codex for software. That’s more interesting to me than another model leaderboard.

I’m still discovering the app, so I won’t pretend to have months of usage data. But I already know why trying it is relatively easy for me: I don’t need to rebuild my workflow inside it. My essential files live in Obsidian — including my identity file, my instructions, my maps, and the skills I use to shape recurring work. I described that architecture in How I built an Obsidian vault no single AI tool can hold hostage.

That’s not a knock on GPT-5.6. It’s a genuinely strong family of models. But models leapfrog each other every few months now, and I’ve stopped treating each new leader as an event. What doesn’t happen every few months is a company restructuring how you’re supposed to work with its AI. That’s what OpenAI actually shipped on July 9 — and it deserves more attention than another scoreboard.

What OpenAI actually shipped

Three things landed at once, and they’re easy to blur together.

First, the GPT-5.6 model family: Sol (the flagship), Terra (the balanced middle), and Luna (the fast, cheap one). The $5/$30, $2.50/$15, and $1/$6 figures are API prices per million input and output tokens — useful context for developers, but not the same thing as the price of a ChatGPT subscription. Second, ChatGPT Work, an agent that takes on long tasks — research, audits, spreadsheets, presentations — and comes back with finished deliverables instead of chat replies. Third, and this is the part most coverage undersold: a new ChatGPT desktop app that bundles Chat, Work, and Codex under one roof.

Here’s the detail that tells you where this is going: the Codex app becomes the new ChatGPT desktop app, while the previous ChatGPT desktop app lives on as “ChatGPT Classic.” Classic is still supported, but the new agent features are concentrated in the new app. When a company calls your current app Classic, it’s telling you something without saying it.

And the reasoning behind the merge is right there in OpenAI’s own numbers: Codex had grown past five million weekly users, and more than a million of them were using it for tasks that had nothing to do with code. Software development turned out to be the test bench for a way of working — give an agent files, tools, and a long-running task, then supervise — that outgrew its original domain. Work is that engine, generalized.

Chat answers, Work delivers, Codex builds

The new app splits your work into three spaces, and the split is smarter than it first looks.

It’s not organized by topic. It’s not organized by difficulty. It’s organized by what you expect back: an immediate answer, a non-technical deliverable, or a software artifact. Ask a question, rephrase a paragraph, think something through — that’s Chat. Need a multi-source research report, an audit, a slide deck, a spreadsheet — that’s Work. Local folders, repositories, terminals, tests — that’s Codex.

This routing happens before the work begins. That sounds like a small interface decision, but it changes the relationship with the tool. I’m no longer choosing only what to ask or which model to use. I’m choosing what kind of work I expect the system to carry.

The routing rule, in one line: instant answer → Chat; deliverable → Work; software → Codex.

I wrote in Claude Code vs Codex that the useful question isn’t which tool is better, but what kind of attention the task needs from you. The new ChatGPT app takes that idea and builds the interface around it. Choosing your workspace is now a decision on par with choosing your model — each space optimizes a regime of work, not a subject area.

That’s the actual news. Not a point on a leaderboard — a product admitting that a chatbot, an agent, and a coding tool are three different working relationships.

Sol, Terra, Luna: pick a level, not a version number

The naming scheme sounds like marketing until you see what it fixes.

Before 5.6, every OpenAI release reset your mental map. With Sol, Terra, and Luna, the number marks the generation while the names mark durable capability tiers that persist across releases. If that sounds familiar, it should — it’s exactly the Opus/Sonnet/Haiku logic from Anthropic. The tier becomes stable vocabulary. You learn “which level for which task” once, and that knowledge survives the next version bump.

Here’s the part I find telling: the most useful comparison isn’t necessarily Sol against the other flagship models. For high-volume, well-defined tasks, the question is simpler: “which is the cheapest model that clears my quality bar?” At $1/$6 per million tokens in the API, Luna may clear that bar for a lot of everyday work — and running the flagship on those tasks means paying for reasoning you never use.

That’s the quiet shift in this release. Even when raw scores are close, competition moves to performance per dollar. The reference chart is no longer only a ranking; it’s a score-versus-cost curve.

The new control panel — and its cognitive tax

Open the model selector in Work and you’ll find three independent dials: the model (Sol, Terra, Luna, plus previous generations), the effort (how long it’s allowed to reason, from low up to max), and the speed. On top of that, Sol gets ultra — a mode that coordinates several agents working in parallel, then reconciles the parts.

The distinction between max and ultra is worth internalizing: max makes one worker think longer; ultra puts a team on the job. A deep but indivisible problem — a counterintuitive bug, an architecture decision — wants max. A broad task that decomposes into independent parts — research, comparison, verification — wants ultra. Paying for more depth doesn’t help a decomposable task, and parallelizing doesn’t help a sequential one.

It’s a genuinely powerful control panel. It’s also a tax. Three dials times several levels is a lot of combinations, and the reflex of “crank everything to the top” can burn tokens without a measurable gain. My practical advice is deliberately conservative: start with Sol at medium effort, then move only when the task gives you a reason. Go up when hidden complexity appears; go down to Terra or Luna when volume matters more than the ceiling.

The skill isn’t knowing the settings. It’s qualifying your task before touching them.

Why I refuse to marry a model

Here’s my broader position, and it predates this launch.

GPT-5.6 leads Claude on coding benchmarks today. The reverse was true a month ago, and it will probably be true again. The leadership flips almost every release now — which means a workflow built around one specific model gets rebuilt every flip, while a workflow where the model is just a setting swaps engines without a rewrite.

I’ve applied this logic to my notes for years — it’s why my Obsidian vault answers to no single AI tool. The durable investment is in what survives the models: your data, your processes, your skills and behavioral contracts, and your evaluation habits.

My workflow lives in files I control. Obsidian holds the notes, Me.md holds the stable context, AGENTS.md points Codex to that context, and my skills remain plain Markdown in the vault. When I switch tools, I may need a different pointer file. I don’t need to rebuild the system underneath it.

That is why I can try the new ChatGPT app without moving into it. The app can become another front door to my work, but it doesn’t become the owner of the work. The tool changes. The workflow stays mine.

So yes, I’m impressed by the new app. And no, I’m not moving in.

Conclusion

Strip away the launch noise and one idea remains: OpenAI stopped selling a smarter answer machine and started selling a set of working relationships — an answerer, a deliverer, a builder — with the model reduced to a dial inside them.

That’s the right direction, and it confirms what this whole year has been teaching: the durable skill isn’t choosing the best model — it’s identifying the kind of work in front of you, then choosing the right workspace and the right amount of reasoning. Models will keep leapfrogging each other. The judgment about what kind of work you’re actually doing stays yours.

Have you tried the new app — and did it actually change how you route your work? I’d love to hear your thoughts.