The cloud agents grew up, and they stayed cloud.

Last Thursday OpenAI shipped the most complete agent product it has ever made. ChatGPT Work runs on the new GPT-5.6 family and connects to more than 1,400 apps. It plans before it acts. It runs scheduled jobs while you sleep. Two days earlier, Anthropic pushed Claude Cowork to mobile and the web, with background agents that keep working after you put your phone away. I read both announcements twice over the weekend, and I want to say the uncomfortable part first: these are good products. The cloud agents grew up.

They also stayed cloud. That is the part I care about, because it is the part that will still matter a year from now. Your context and your finished work now live in a vendor workspace, on a vendor meter. This post is my read of what shipped and of the trade both vendors are asking you to make.

What shipped last week

ChatGPT Work is the clearest statement OpenAI has made about where it thinks knowledge work is going. The engine is GPT-5.6, which arrives in three sizes:

• Sol, the flagship, at $5 per million input tokens and $30 per million output.
• Terra, the balanced middle, at $2.50 and $15.
• Luna, the fast one, at $1 and $6.

Around the models sits the product. More than 1,400 connectable apps. A Plan mode that drafts an approach and shows it to you before executing. Scheduled Tasks that run without you. And Codex folds into the ChatGPT desktop app, so the coding agent and the chat agent finally share one surface.

Anthropic's week was quieter but pointed. Claude Cowork began in January as a desktop research preview: an agent for non-technical knowledge work that operates directly on files in folders you choose, built on the same agentic core as Claude Code. On July 7 it expanded to mobile and the web, with background agents that keep a job running after you close the tab. Underneath it sits Claude Fable 5, generally available since June, with a million-token context window priced at $10 per million input tokens and $50 per million output.

What the vendors got right

I run a company that competes with these products, so discount my opinion accordingly. Here is what I think they got right.

Plan mode is the right idea. Cursor 3 shipped a Plan Mode in April, and Anthropic's Ultraplan landed the same month. Now ChatGPT has one. When that many serious teams converge independently, the lesson underneath is real: an agent that drafts its approach for review before executing produces better work than one that starts typing immediately. We agree so strongly that OMEGA's Project Canvas has been plan-as-nodes from the start. A plan you can see is a plan you can fix before it costs you anything.

The connector count is not a vanity number. 1,400+ apps means the agent can reach the systems where work already lives, and Viktor, the Slack and Teams AI coworker, raised a Series A in May largely on the strength of 3,200+ managed connectors. Investors think integration breadth is the moat, and for a team that lives inside SaaS tools all day, an agent that reaches those tools without setup is useful on day one.

And background agents solve a real problem. An agent that dies when your laptop lid closes is a toy for anyone whose calendar has meetings on it. Cowork on mobile and the web means you can hand off a task at your desk and check the result from a hallway. That is the correct user experience, and pretending otherwise would be marketing.

Where the work lives now

None of that changes the shape of the deal. The agents grew up inside the vendor's house, and your work moved in with them.

Trace a single task through ChatGPT Work. The connectors pipe your email and documents through the vendor's infrastructure. The plan is drafted there, and the output is written there. The accumulated memory of the task, the context that makes the agent smarter about you next month, accrues to an account you rent. Every token is metered by the party that also sets the price of a token. Each step is reasonable. The sum is a workspace you do not control holding work you cannot easily take with you.

Portability is the tell. Try to move two years of accumulated context out of ChatGPT's memory and into Claude, or the reverse. It stays behind. That was true of every platform before this one, and agents raise the stakes, because agents accumulate far more than chat history ever did: your files, your standing tasks, your approval habits, the whole texture of how you work.

The failure modes are ordinary business, not villainy. A hosted workspace can be repriced, and when the meter and the workspace belong to the same vendor there is no routing around it. Models get retired too, which shifts the behavior your scheduled tasks quietly depended on. And an export that technically contains your data can practically preserve none of what made the agent good at your job. You do not need to believe anyone is acting in bad faith. You only need to notice who holds which keys.

Cowork's own trajectory illustrates the pull. The January preview was, to me, the most interesting agent release of the year, precisely because it worked on files in folders on your own machine. The July expansion goes to mobile and the web, and an agent that keeps working after the tab closes is by definition not running on your hardware. I understand the reasoning; a background agent cannot live on a phone in a pocket. But watch what moved. The agent left your folders and went home to the cloud.

Both camps keep reaching for the desktop

Here is the detail that convinces me location is the real question: the traffic runs in both directions.

Manus, the archetypal cloud agent and a Meta property since December, shipped a desktop app called My Computer back in March. Anthropic started Cowork in your local folders before growing it toward the cloud. Meanwhile the local-first side keeps maturing: OpenClaw is an open-source, local-first agent with its ClawHub skill registry and a menu bar presence, and Hermes Agent, Nous Research's self-hostable open agent, shipped a native desktop app in June. Everyone is circling the same fact from different directions: the valuable context, the files and the daily work, lives on your machine. The cloud vendors want a bridge to it, and the local tools want to meet you there.

So the framing of last week comes down to which direction the bridge runs: your data commuting to the vendor's workspace, or the intelligence commuting to yours.

The same jobs, run from your Mac

OMEGA's answer is a different location for the work. OMEGA is a native macOS app, macOS 14 and later on Apple Silicon, and the workspace is your machine.

The concrete version. Take a task I would happily hand to ChatGPT Work: read a folder of client contracts, pull renewal dates and obligations into a summary, produce a client-ready document, and set the job to re-run monthly as contracts change.

In OMEGA, the folder never uploads by default. Project Canvas imports files, local folders, git repos, and Figma links directly, and the agent reads the contracts on disk. The model doing the reading is whichever one you chose for the job: a local model through Omega-MLX on Apple Silicon when the material is sensitive, or a frontier model through your own keys when you want the horsepower. BYOK means the provider's price passes through with no markup on tokens, and if you already pay for a Claude or Codex subscription, OMEGA can route through that instead of charging you twice for intelligence you already bought. The finished document exports in any of 47 formats with your Brand Kit applied, and it lands on your disk rather than in a cloud drive behind a share link.

The monthly re-run is an Automation: you describe it in plain English, the builder compiles it into triggers and schedules, and when it needs to run while your Mac sleeps, an always-on VPS runner executes it. I will be straight about that last clause, since this whole post is about where things run. The runner is a server. The difference is what lives on it: the schedule and the standing instructions, not your accumulated context and not your files. If you already have n8n graphs, they import directly.

Memory is the sharpest contrast of all. OMEGA's four-layer memory and Brain knowledge graph accumulate on your machine, and they are provider-portable. Swap models or providers mid-chat and the context comes with you. Our standard engine, Neural-Fractal Agentic AI™ (NFA), decomposes big jobs into small scoped agents drawn from a catalog of more than 200, carrying 5,300+ bundled skills, and every one of them reads and writes that same local memory. When a better model ships next month, and lately one ships roughly monthly, you point your existing workspace at it the same afternoon, memory intact.

The rest of the surface follows the same rule. Computer use asks per-app approval before it touches anything, and voice runs on-device. Budgets and per-workspace keys make the meter one you configure rather than one you are subject to.

The trade-offs, stated plainly

The cloud agents beat us on real things, and I would rather you hear the list from me.

• Zero setup. ChatGPT Work is a login and Cowork on the web is a URL. OMEGA is a download, and a native app asks more of you up front than a browser tab does.
• Away-from-desk work. Background agents in someone else's cloud genuinely keep working while your machine is off. Our answer for standing jobs is the Automation runner; interactive agent work wants your Mac awake.
• Team-wide rollout. If your whole company already lives in a vendor workspace, an agent inside it inherits your permissions and sharing on day one. A local-first tool has to earn its way onto each machine separately.
• Hardware. OMEGA requires Apple Silicon. That requirement buys fast local models through Omega-MLX, and it costs us everyone on a PC.

If those cut against you, use the cloud products. They are good, and I would rather lose a reader than win one with a comparison that hides the ball.

Same models, different landlord

The thought I keep returning to is that the intelligence in every product named in this post is shared substrate. Fable 5 is on the API. The GPT-5.6 family is on the API. The weights ChatGPT Work calls and the weights a BYOK key reaches are the same weights. What the July launches actually shipped, the thing the vendors are actually selling, is the workspace around the weights: the connectors, the memory, the file store, the meter.

That workspace is the product, and it is also the lock. My bet with OMEGA is the opposite arrangement: the workspace should be the one thing you own outright, and the models should be the interchangeable part you rent at cost.

The cloud agents grew up last week, and I mean that as a compliment. But growing up means deciding where you live. They chose the vendor's house. We built OMEGA so you could choose your own.

If you want the two approaches side by side, our comparison page goes feature by feature, and pricing is public. The longer version of this argument is in why we built OMEGA.