Viveka 
Written in July 2026. These tools move fast, so treat the product specifics below as a snapshot — the way to choose between the modes is the part that lasts.
What’s the difference between ChatGPT Work and Claude Cowork?
Both are agent modes: instead of answering a question, they take on a whole job. You hand one an outcome, it pulls context from the apps and files you’ve connected, writes a plan, works through that plan step by step for as long as hours, stops to check with you at the big decisions, and returns a finished file — a spreadsheet, a deck, a document, or a working web app you can send someone a link to.
If you already use Claude Cowork, ChatGPT Work will feel familiar — same idea, newer build. The honest difference right now, at least in my testing, is integrations: Claude connects to more of your apps and files and has a bigger library of skills and connectors to pull from, and ChatGPT’s version is strong and closing fast. Both also have “skills” now, which are reusable instructions you build once and run over and over, though I find Claude’s easier to make.
Chat, Work, or Code — which mode should you actually use?
Use Chat to think out loud, Work or Cowork to hand over a job and walk away, and Code or Codex for anything that ends up as working software. Both apps now have all three: ChatGPT has Chat, Work, and Codex; Claude has Chat, Cowork, and Code.
Chat is your questions and explanations, the twenty minutes you spend turning a mess in your head into a plan you can act on. Work and Cowork are for the folder of client notes that needs to become one report, the two hours of research you want back structured, the deck you’ve been avoiding for a month. Code and Codex are for the things you own at the end — a site, a prototype, a dashboard, a tool. I built my VIOS app and its landing page entirely in Claude Code, which a year ago I’d have had to hire out (it’s the system I now take women through in about two hours).
Why does the mode matter more than the model?
Because the mode decides whether you get a plan, a finished deliverable, or working software — and almost everyone only ever opens the first one. They type a question into Chat, get a mediocre answer, and decide the whole category is a little oversold, which suuuucks, because those are usually the exact people who’d get the most out of the other two.
If you’ve been living in Chat and wondering why this all feels like a slightly better search engine, that’s your answer. The prompt you write matters. The mode you open matters more. I wrote a whole piece on choosing tools that amplify your voice instead of eating your week, and this is the same idea one level up: pick the right room before you worry about the right words.
Which is smarter right now — ChatGPT’s Sol or Claude’s Fable?
On the public benchmarks they land within about a point of each other, with wildly different price tags, which means the fight has moved from capability to billing — what these two are really competing over now is your monthly bill. So “which is smarter” is mostly the wrong question.
According to the public benchmarks, as opposed to my own one-off tests, ChatGPT’s Sol scored within about a point of Claude Fable 5 on the Artificial Analysis Intelligence Index while costing something like a third to a half as much per task, and it took the Coding Agent Index outright; Fable still wins the hardest evaluations, including a roughly fifteen-point gap on SWE-bench Pro and the lead on the tests built to imitate real knowledge work. (Yeah — I have no idea what half of that means either.) When I split-tested both on a deep audit of my own websites, Fable’s results came back better because of the connectors, but Sol is cheaper and within reach of a lot more people. Two very good models, about a point apart, priced worlds apart.
Why does AI voice mode get simple things wrong, like counting letters?
Because these models read words in chunks, not letter by letter. Ask a voice assistant how many E’s are in “seventeen” and it will answer wrong without blinking; ask the typed version the same thing and it gets it right instantly.
The word “seventeen” is one or two blocks to the model, not nine characters, so counting the letters inside it means slowing down and pulling the word apart. The typed version has time for that. The voice version is built to answer in a couple hundred milliseconds, because a two-second pause would wreck the conversational feel, so the slow careful path is exactly the one it skips. Once that clicks, a LOT of otherwise-weird AI behavior starts to make sense. (The newest voice modes are genuinely impressive, by the way — they listen and talk at the same time now, so you can cut them off, change your mind mid-thought, and they keep up. Creepy and cool in equal measure.)
What else changed — and the part almost nobody covered
Alongside the launch, OpenAI shipped shareable interactive pages, scheduled tasks that run on their own, and folded its separate coding and browser tools into one desktop app. The piece the coverage skipped almost entirely: the model’s release was staggered at the request of the federal government.
Axios reported it, and Sam Altman confirmed to CNBC that OpenAI made “many changes” after a back-and-forth with the administration, with the model going wide only after the Commerce Department’s Center for AI Standards and Innovation finished evaluating it. A frontier AI release is now something a federal agency signs off on before you’re allowed to touch it, and it got buried under a demo of a landing page somebody built from a two-sentence prompt. (Cool cool cool.) Meanwhile the two companies spent launch week resetting each other’s usage limits in public, which was petty and delightful and, briefly, a win for the rest of us.
What should you actually do with any of this?
Stop opening Chat for everything. This week, take one job you’ve been putting off — the messy folder, the pile of client notes that needs to become a single report, the deck you keep not making — and hand it to Work or Cowork instead of asking Chat a question about it.
Give it a real task, not the one every YouTube video shows you (about 90% of which are “clean up my email” — not useless, but I bet you can find something better). The first run might take longer than doing it yourself would have, and it will show you exactly what you got wrong in how you asked, which is what the first run is FOR. It gets faster after that, and then it starts giving you time back. Point one of them at your nightmare of a Desktop or Drive and let it organize the whole thing (mine felt like it got a colon cleanse afterward — squeaky clean), or have it run a deep audit of your website for consistency, SEO, branding, voice, and broken links, and then actually go fix what it finds.
So which one should you pick?
Pick on integrations and price, not hype — and honestly, if you can swing both, use each for what it’s good at. But here’s the part that matters more to me than any launch: everything above is a story about two enormous companies competing over who gets to do your work for you, cheaper, and not one word of it is about what only you know.
The models are a point apart and getting cheaper every month, so the one thing here that won’t get cheap is the thirty years of judgment you’re walking around with. Use the agents. Use both if you can. And then go put your name on the part only you could have written.