OpenAI shipped ChatGPT Work on July 9, 2026, and the first thing most people got wrong about it is the mental model. This isn’t a smarter chatbot. It’s an agent that takes a goal, disappears for a while, and comes back with a finished spreadsheet, a slide deck, or a working web app. You don’t watch it think. You wait, and then you get a deliverable.
That sounds great until the bill arrives. Because Work doesn’t run on the flat “you pay $20 and use it all month” logic that made ChatGPT Plus feel safe. It draws from the same credit pool as Codex. Let an agent grind on a research task for an hour and you’re spending credits the whole time — and unlike a chat message, you can’t eyeball how expensive a run will be before you fire it.
So the real question isn’t “is ChatGPT Work good.” It usually is. The question is whether the usage-based bill is worth it for the kind of work you do. The numbers below settle it either way.
What ChatGPT Work actually is
OpenAI now splits the ChatGPT desktop app into three lanes, and understanding the split is the whole game.
Chat is the thing you already know — quick back-and-forth, questions, drafting, brainstorming. Codex is the coding agent, built for long-horizon engineering work. Work sits in the middle: an outcome-oriented lane for research, analysis, and finished business documents. You hand it something like “pull our last four quarters of churn, find the three biggest drivers, and build me a slide deck,” and it gathers context across your connected apps, breaks the goal into steps, and returns the deck.
The Codex app is merging into this new unified desktop app, so Chat, Work, and Codex now live side by side on every plan — including Free. That consolidation is the strategic story: OpenAI watched its own engineers move 85%+ of their output to Codex-style agentic work and decided the whole company — legal, finance, recruiting — should run that way too. Work is how they’re selling that shift to everyone else.
Here’s the part that matters for your wallet: Chat is mostly covered by your subscription. Work is metered.
The billing trap, in plain numbers
Workspace agents ran free during a research preview that OpenAI kept extending. That free ride ended July 6, 2026. Since then, agent runs invoked inside ChatGPT draw down credits.
Each run is metered across three separate token streams — regular input tokens, cached input tokens, and output tokens — each with its own credits-per-million rate. You don’t need to memorize the rate card. You need to internalize the ranges, because they’re wide:
- A typical end-to-end agent run on GPT-5.5 lands somewhere around 5 to 25 credits.
- A ChatGPT for Excel/Sheets task runs roughly 5 to 20 credits.
- A PowerPoint task runs 10 to 50 credits.
Look at that PowerPoint spread. A ten-slide deck and a fifty-credit deck can be the same request — the cost depends on how much the agent reads, retries, and reasons along the way. That’s the trap. With chat, a bad answer costs you nothing but a re-prompt. With an agent, a run that wanders costs credits whether or not the output is any good.
One thing that softens it: credits only apply to advanced models and agentic features. Core stuff — search, file upload, canvas, the everyday chat models — doesn’t touch your credit balance. So you’re not bleeding credits just by opening the app. You’re spending them specifically when you send an agent off to do multi-step work. That’s a cleaner line than a lot of people assume, and it means the discipline is simple: know when you’re invoking an agent versus just chatting.
What each plan realistically buys
The plan ladder as of July 2026 runs Free, Go, Plus ($20), Pro (offered at $100 and $200 usage tiers), Business (about $20 per seat annual, $25 monthly), and Enterprise (custom).
The catch is that not everything in Work is available everywhere. Ultra mode in Work — the heaviest, most thorough agent setting — is Pro and Enterprise only. Workspace agents and company knowledge, the features that let an agent pull from your org’s connected data, are Business and Enterprise only. So if your reason for wanting Work is “have it read our internal docs and build reports,” Plus won’t get you there no matter how many credits you throw at it. You need Business at minimum.
For Business plans, Work, Workspace Agents, ChatGPT for Excel, and ChatGPT for PowerPoint all draw from one shared agentic credit pool. That’s convenient for billing and dangerous for budgeting — one team member running deck-generation all afternoon eats into the same pool everyone else needs.
My read on the tiers:
- Plus ($20) — fine for occasional Work runs on your own files. No company knowledge, no workspace agents. Treat it as “agent for personal deliverables,” not “agent for my team’s data.”
- Pro ($100 / $200) — you get Ultra mode and a much bigger usage allowance. This is for people who run agents daily and would blow past Plus limits by Tuesday.
- Business ($20–25/seat) — the real unlock for teams, because it’s the entry point for workspace agents and company knowledge. Watch the shared credit pool.
- Enterprise — custom everything, plus the admin controls to actually cap runaway spend.
Where Work earns its credits — and where it wastes them
Work shines on tasks where the deliverable is worth more than the credits it burns. Multi-source research where you’d otherwise open fifteen tabs. Turning a messy dataset into a clean analysis. Building a first-draft deck or a structured doc from scattered notes. Scheduled Tasks that run a report every Monday without you touching it. In all of these, you’re trading real human minutes — sometimes hours — for a handful of credits. Easy math.
Where it wastes credits is the stuff you reflexively throw at it out of habit. “Summarize this email.” “Rewrite this paragraph.” “What’s the capital of Portugal.” Those belong in Chat, which won’t touch your credit balance. Sending a quick one-off to the agent lane is like hiring a contractor to change a lightbulb — it’ll get done, but you’re paying agent overhead for a chat-sized job.
The habit worth building is this: before you invoke Work, ask whether the task has steps. If it’s one action, use Chat. If it’s “go do this sequence of things and bring me the result,” that’s what Work is for and what the credits are buying.
ChatGPT Work vs the other outcome agents
Work isn’t alone. By mid-2026 every major lab shipped its own “agent that finishes the job,” and they’ve split along a genuinely useful line: where the agent lives.
Claude Cowork is device-attached and file-centric — an autonomous desktop agent that runs multi-step work unattended across your files and connectors, included in Claude Pro at $20/month. It’s the one people reach for when the deliverable is an Excel model, a Word doc, or a PDF, because it’s sitting right next to your files.
Gemini Spark is identity-attached to your Google account rather than a device, so it keeps working in the background regardless of what your laptop is doing. It’s the pick if you live in Gmail and Google Workspace and want inbox automation and 24/7 background tasks. It rides on Google AI Ultra at $100/month.
Microsoft Copilot is brilliant inside Word, Excel, and Outlook, but its true agent layer is a separate product — Copilot Cowork went generally available in June 2026, and full agent capability runs roughly $30 plus $15 per seat once you add Agent 365. Notably, Microsoft’s agent layer runs on Anthropic’s Claude models under the hood.
Where does that leave ChatGPT Work? It’s the outcome agent for people already anchored to OpenAI — the ones who want research, analysis, and finished files inside the same app where they chat and code. Its edge is the three-lanes-in-one-app consolidation. Its weakness is the pricing model: Cowork gives you agent runs inside a flat $20 subscription, while Work meters you. If your usage is spiky and low, Work’s pay-for-what-you-run model can actually be cheaper. If your usage is heavy and constant, a flat-rate competitor may win on predictability alone.
I’ll say the quiet part: for a lot of solo users, Claude Cowork’s flat pricing removes exactly the anxiety that makes Work annoying. You don’t do mental math before every run. That peace of mind is worth something.
A decision guide by who you are
Solo operator or freelancer. Start on Plus at $20 and be honest about your volume. If you’re firing a couple of Work runs a week to build client deliverables, the credits are trivial and you’re fine. If you notice yourself reaching for the agent lane constantly, either move up to Pro or seriously price out Claude Cowork’s flat plan — because at high volume, unpredictable per-run billing stops being a feature.
Small team. You almost certainly want Business, not a pile of Plus seats — that’s the only way to get workspace agents and company knowledge, which is the whole reason a team wants an agent reading its shared data. Budget for the shared credit pool and set expectations early. One person left running PowerPoint generation all day can quietly drain what the rest of the team needs.
Enterprise. The usage model is a feature here, not a bug, if you use the admin controls. Set per-seat or per-department caps, watch the dashboards, and route heavy recurring work through Scheduled Tasks so it’s predictable. The failure mode is handing 200 people an uncapped agent and discovering the bill at the end of the month. Cap first, scale second.
Across all three, the same rule caps runaway spend: keep Chat and Work separate in your head. Chat is the free-flowing, subscription-covered lane. Work is the metered lane you deploy on purpose. The people who get burned are the ones who treat every prompt like an agent job. The people who come out ahead treat the agent like what it is — a contractor you send off with a clear scope, not a chat buddy you poke all day.
If you want to see how OpenAI’s model tiering feeds into all this, the GPT-5.6 tier routing breakdown covers which model each lane leans on and why that changes your per-task cost — the same reasoning that decides a 5-credit run from a 25-credit one.
One thing to try this week: pick a task you’d normally grind through by hand — a competitor scan, a monthly report, a deck from raw notes — and run it once through Work. Note the credits it burned. That single number tells you more about whether Work belongs in your budget than any pricing table can.
Sources: OpenAI Help Center — ChatGPT Work and Codex, OpenAI Help Center — Codex rate card, OpenAI Help Center — Using Credits for Flexible Usage in ChatGPT, The Decoder — OpenAI pairs GPT-5.6 rollout with ChatGPT Work, TechTimes — Workspace Agents credit pricing, Cowork Certification Guide — Cowork vs ChatGPT, Gemini, Copilot. Pricing and credit ranges are as of July 2026; check OpenAI’s rate card for current figures.