OpenAI has the smarter model. Anthropic is winning anyway.
Anthropic figured out something OpenAI didn't — and it has nothing to do with the model.
Two charts tell the story.
On Artificial Analysis’s intelligence ranking, OpenAI’s frontier models still beat Claude. They’re ahead by the benchmarks that have defined this industry for three years.
On Ramp’s AI Index for May 2026, Anthropic just overtook OpenAI in enterprise spending.
Anthropic usage has been exploding, growing past OpenAI for the first time this month.
If intelligence is what matters, this shouldn’t happen.
So what is Anthropic doing that OpenAI isn’t?
OpenAI is competing at the model layer. Anthropic is competing at the platform layer — and that’s where applied AI gets won.
Two bets got Anthropic there. Neither shows up on a benchmark.
Bet #1: Mind and hands
MIT’s motto is mens et manus. Mind and hands. The idea: intelligence is meaningless without the ability to act on the world. You can be a structural engineering savant, but without a saw you’re not building a house.
For three years, the AI industry has been obsessed with the mind. Bigger models. Smarter benchmarks. More parameters.
Anthropic bet on the hands.
It started with MCP — the Model Context Protocol. A standard that lets any AI model access any tool, any data source, any system. Not glamorous. Not a benchmark-topping release. Just “plumbing”.
Then came Claude Code. The dirty secret of Claude Code’s success: it didn’t win because the underlying coding model was better. The same Claude models were available in Cursor, in every IDE plugin, in every wrapper on the market. Claude Code won because it lives in the terminal — where thousands of tools already exist. Git. Docker. ffmpeg. jq. Every CI script ever written.
Putting Claude in the CLI didn’t make it smarter. It gave it hands.
The pattern repeats with Claude (skills), Claude Cowork (file access), and Claude Design (HTML rendering). None of these are “smarter Claude.” They’re Claude with different hands — different tools, different surfaces, different ways to act in the world.
OpenAI spent the same period chasing the mind. New frontier models. Better image generation. Reasoning upgrades. The benchmarks went up. The enterprise spend went sideways.
Bet #2: Platform > applications
The second thing Anthropic got right: it has focused on building a platform, not just applications.
This is the difference between Microsoft in the 90s and Meta in the 2010s.
Microsoft built Windows (the platform) and Office (the killer app). It was its own first and best customer — but never its only customer. Microsoft execs were emphatic about creating value for developers… developers, developers, developers. The result? A generational platform that has thrived through every era — even the ones it didn’t win.
Meta took a different approach. Forced to live on platforms it didn’t own — iOS, Android, Mac, and Windows — Meta focused on building the most engaging applications on those platforms. As a developer platform, they were reluctant. F8 was always more about Meta’s product launches than developer tools.
Meta executed their application strategy brilliantly but also out of necessity because others had the platform locked up. A lot of OpenAI’s senior leadership came from Meta. And it shows.
ChatGPT has been largely closed to developers. OpenAI’s strategy: win the consumer app, charge a subscription, sell ads later.
It’s the Meta playbook.
Anthropic is running the Microsoft playbook. Claude is open. MCP is open. Skills are open. Bring your own tools, your own data, your own integrations. Build on top. The interfaces are there. And Anthropic uses its own apps — Claude Code, Claude, Claude Design — the way Microsoft used Office: as proof the platform works, not as the only thing customers are allowed to run on it.
AI’s next era
Benchmarks aren’t irrelevant. A significantly less capable model would have lost regardless of the platform it’s plugged into. Hitting benchmarks is necessary, not sufficient.
As AI models have matured, the gradient has shifted from the marginal IQ point to the harnesses, tools, and workflows that intelligence plugs into.
Enterprise software has always rewarded platforms over applications: SAP, Salesforce, AWS, Microsoft. Anthropic recognized this early, shifting focus away from intelligence benchmarks to real-world impact.
Now that we’re in the platform era of AI, the most important question for your company isn’t which model is most capable. It’s:
“Which model best orchestrates the agents, tools, workflows, and processes that run my business?”
The thing you’re pointing at probably needs a name. Not the model, not the app — the layer in between. What tools the model sees, in what order, what each tool’s description tells it about when to reach. Where the harness ends and the agent begins.
I keep running into this at work and there’s no good word for it. “Prompt engineering” is way too small now. “Agent design” makes it sound more sci-fi than it is — most of the job is writing tool descriptions and arguing about which five things should be in the default context. It’s plumbing work.
My half-joke bet: harness PM in 2026 is what growth PM was in 2014. Nobody had the title, then suddenly everyone needed one, then it got boring and became part of the job.
I disagree on the platform side. Anthropic tries to compete with everyone that starts making money using their model. That's not a platform play at all. Platforms have very clear rules on what the platform has to own (don't build on top of us for this), what the platform competes on (we'll build in this space too, but we don't prevent you from building something better), what the platform wants people to build that it won't, and what it won't build and doesn't want other people to build. Anthropic hasn't been clear about this at all. Because of how expensive they are to run, if we make money on top of them, they want to take it. That's a terrible developer brand long term.