OpenAI, Microsoft, and Anthropic all bet on one app to run your workday

The big AI players just made the same bet at the same time. OpenAI shipped ChatGPT Work. Microsoft is building a Copilot "super app." Anthropic is folding its Cowork tool into Claude. They all want one app that does your knowledge work, not a chatbot you poke at. Let me catch you up on what changed and what it means for the thing you ship.

One app to eat the workday

The pitch is simple. Stop bouncing between a chatbot, a coding tool, and an agent. Put it all in one place. Microsoft CEO Satya Nadella told investors the company will bring its Copilot experiences "together, including Code, in one 'super app'" by the end of summer, per Mashable's rundown. OpenAI already pushed its version to employees. Anthropic is doing the same with Claude.

There's a land grab underneath the framing. When you can call InstaCart, Uber, Zillow, and OpenTable from inside ChatGPT, that app absorbs traffic that used to be yours. The goal is to own the whole session, not one task inside it.

OpenAI's ChatGPT Work crossed 10 million users in three weeks, and Chat and Work merge by year end. So this is not a power-user side product. It's a preview of how a billion weekly users will work.

The agent that grinds while you sleep

The demos are the part worth watching. On Lenny's how-I-AI episode, OpenAI's Nick Baumann feeds 50 or 60 raw clips into a plugin, talks through the story he wants, and the tool transcribes everything, picks the best takes, and assembles finished vertical videos overnight. A content team's whole day, handled while he sleeps.

He also books a flight, files an expense report, and checks his calendar in one voice conversation, without opening a single app by hand. The agent runs for hours on a cloud computer of its own, and "Heartbeats" let it wake itself to flag a suspicious charge or a nearby package. No code required to set that up.

The lesson for your product: users are starting to hand over full jobs, not single steps. A tool that only answers questions will feel thin next to one that finishes the work.

Talking beats typing

One quiet product insight from the Codex walkthrough: voice gets people to give more context. Claire Vo calls it the "yapper's API." People freeze at an empty prompt box and try to be precise. When they talk, they ramble, and the extra context makes the output better.

Speed is turning into a real design call too. Baumann uses the highest-intelligence mode in Voice, but once a model has the tools it needs, latency starts to matter more than another small bump in reasoning. Tools built around the idea that waiting is fine will lose to tools that feel instant.

Both are cheap moves you can steal. Add a voice or free-text on-ramp. Cut the wait where you can.

Apps didn't die, they multiplied

The counter-story matters for your roadmap. Even as agents try to swallow the app layer, new app releases jumped 60% year over year worldwide in Q1 2026, and 80% on iOS alone. AI coding made shipping cheap, so more people ship.

Meta is running the same play at scale. On its earnings call, Mark Zuckerberg said LLMs make it "a lot easier to ship new apps", and the company is spinning out standalone apps for Groups, Marketplace sellers, and more. Threads finally hit 500 million monthly users, partly on LLM-powered recommendations.

So two things are true at once. The majors want one super app. And the cost of building a sharp single-purpose app keeps falling. Both are real. Pick which game you're in.

The deep cut

Look at how OpenAI built Work, not just what it does. The agent gets broad freedom inside a task, but continuity across tasks runs through a controlled product layer: Library, Projects, Personal Context, Memory. The agent can't rewrite that layer at will. Shlok Khemani's teardown found even a small crack, where one thread keeps reading a stale local copy of a file after another thread changes the canonical version.

That's your opening. The hard part of agent products is not the flashy single task. It's state, memory, and file truth across many tasks over time. Even OpenAI hasn't fully solved it, and there's no meta-layer agent coordinating between tasks yet. If your team owns a narrow domain, that boring plumbing is where you can be clearly better than a generalist super app.

Three questions for your team

1. If a user could hand our product a whole job instead of one step, what job would that be, and what's stopping us from finishing it end to end today?
2. Where in our flow do we force people to type precise input when a voice or loose free-text on-ramp would get us more context and a better result?
3. When a user works across many sessions, does our product keep their files, memory, and state straight, and can we prove it, or do we have a stale-copy problem waiting to surface?