Greg Brockman tweeted something tonight that landed: "the model alone is no longer the product."
Eight words. Almost 877,000 views in a day. I had two reactions at once. The first was full agreement, the kind you have when somebody finally says out loud the thing you've been saying for a year. The second was surprise that I'm not seeing this framing everywhere already, because the evidence has been piling up for months.
This is something we've been writing about, talking about with clients, and building our consulting practice around. So rather than just nod at the tweet and move on, here's where we've been saying it, both publicly and in client work.
What we've been publishing
The clearest version of this argument from us went up on May 4, 2026, in a piece called Anthropic and OpenAI Are Building Implementation Arms. That's a Good Thing.. The framing was almost identical to Brockman's:
"The bottleneck is not the model. It's everything around the model."
We wrote that piece on the day Anthropic and OpenAI both announced new implementation arms, a $1.5B joint venture and a $10B vehicle respectively, designed to embed engineers inside client organizations. Two of the most valuable AI companies in the world were telling the market, in the most public way they could, that the technology by itself is not enough.
A few weeks earlier, in The Enterprise "Brain" Is an Architecture Problem, we put it in terms of why coding assistants feel so much further along than knowledge work assistants:
"The reason coding assistants feel magical is not the model. It is the container."
Code lives in a repo. Knowledge work lives in fifteen Slack threads, six Notion pages, three SharePoint folders, and the heads of two people. The model can be the most capable system available and it still has nothing to read.
In Skills Don't Sync Yet, we took the framing further and described what's actually being assembled around the model:
"The connectors are the device drivers. The skills are the applications. The plugin system is the app store. The agent itself is the kernel."
That's not a chatbot. That's the early shape of an operating system for knowledge work. And once you see it that way, asking "which model is best" starts to feel like asking which kernel is best. The honest answer is that it matters less than people think, and almost nothing else about the experience flows from that choice.
Cowork Is More Than Skills. It's Modular Context. was the piece where we tried to explain why two AI tools running the same underlying capability can feel completely different to a knowledge worker. What separates them is how context accumulates, persists, and compounds across sessions, files, skills, and connectors. The tool that does that well will outperform the tool that doesn't, even when their model benchmarks are within a hair of each other.
AI Skills Are Infrastructure Now. Who's Managing Yours? made the operational version of the argument. Once you accept that the value lives in the ecosystem around the model, you have to accept that someone in the organization needs to own that ecosystem. Skills, context, connectors, governance, and cost control all need an owner. It's infrastructure, and managed infrastructure needs a team.
A Pattern Is Emerging: Most Companies Need Three AI Platforms was the procurement version. If the model itself is no longer the differentiator, the right portfolio looks different than it did a year ago. The framework we laid out was the one you already have through your office suite, the one you invest in heavily, and the one you keep around for optionality. That only works as a recommendation in a world where model choice is one input among many, not the whole game.
And in Execution Is Solved. Now What?, we made the personal version of the same point. A single consultant can now run four agent sessions in parallel, ship client deliverables, and publish content at this pace because the layer above the model finally got rich enough to compose real work out of it.
Across six articles, the underlying claim is the same. The model isn't the thing the way it used to be.
What's been coming up in client meetings
The same conversation has been showing up across client engagements for months. A few patterns worth naming, without identifying anyone.
In one weekly AI working session a few weeks ago, the team was evaluating a vendor platform that aggregates access to multiple frontier models for a flat fee. The instinct in the room was that this would solve their AI strategy. We pushed back on it. The clearest line from the meeting summary captured the conclusion well: participants emphasized the importance of integrated ecosystems for AI tools rather than simple model aggregation. Paying for access to multiple frontier models, without the surrounding ecosystem of skills, connectors, context, and workflow, doesn't actually get you anywhere on its own.
In another weekly session, the question was whether to keep building a custom AI solution or transition to an enterprise platform. The conversation was framed around features, integrations, and how easily the platform plugged into the tools the team already used. Nobody at the table was arguing about benchmarks. They were weighing ecosystem fit.
In a corporate education session on context engineering, the entire workshop centered on the ecosystem around the model rather than the model itself. We walked through personalization, memory, the library for stored files, custom GPTs, projects, and skills. The point we kept hitting was that the difference between a frustrated user and a productive one almost never comes down to which model they are talking to. It comes down to how well they have set up the environment around that model.
A separate engagement made the same point from the Claude side. The deliverable for that client was a custom Claude Cowork plugin: a set of modular skills bundled together, an MCP integration into their existing systems, and scheduled tasks that ran key workflows on their behalf. The whole effort was about building the layer around the model so it could actually do useful work for them.
There was also a conversation last week that captures the gap we keep running into.
I made the comment that Claude is hands down the enterprise leader right now. The response that came back was, "Oh, it's even better than ChatGPT 5.5?"
That right there is the misunderstanding. The question assumes "leader" means "the best model at the math problem." It doesn't. The frontier models from the major labs are all extraordinarily capable for the vast majority of knowledge work that needs to get done. They all do adaptive model selection and reason when reasoning is warranted. For deep engineering and novel science, there is still meaningful differentiation between models, and that conversation matters in those contexts. For the rest of knowledge work, the differentiation almost entirely lives somewhere else.
It lives in how the tool handles files on your machine, whether your skills follow you across surfaces, how connectors and scheduled tasks work, how context persists, and how artifacts get created and reused. It lives in the developer ecosystem, the rate at which new capabilities reach the user, and whether the whole thing feels like an operating system or a chat window.
When we say Claude is the enterprise leader right now, that's what we mean. Even after ChatGPT's introduction of skills and agents, the Claude ecosystem just works better for real knowledge work today. The gap may close, and we watch it weekly. But the conversation needs to be about the ecosystem, not about model version numbers.
Where this leaves us
Brockman's tweet was eight words, and it's getting attention for a reason. We're glad to see the framing move into the mainstream, because the organizations that internalize it now will be ahead of the organizations that are still litigating which model is fractionally better at a coding eval.
The model alone is not the product. It hasn't been for a while.
Tweet that spurred this piece: Greg Brockman, May 21, 2026