Today's signal
Cursor just launched Composer 2.5, its most capable coding model yet. It scores 79.8% on SWE-Bench Multilingual, nearly matches Anthropic's Opus 4.7 across most evals, and is priced at $0.50 per million input tokens. That last number is not a typo.
But the product launch is not the real story.
The real story is what Cursor has quietly become. Two years ago, it was a VS Code fork that called the OpenAI API. Today, it is training its own frontier models on open-source checkpoints, running that training on xAI's Colossus 2 supercluster (one million H100-equivalents), and shipping models that compete with the most powerful labs in the world at a fraction of the API cost.
Why it matters
The standard playbook for AI application companies has been: pick a foundation model provider, build a product layer on top, and hope your UX moat outlasts the next API update. That playbook has a fundamental ceiling. When your core intelligence is a commodity rented from OpenAI or Anthropic, your margin, your differentiation, and your roadmap are all hostage to someone else's decisions.
Cursor just blew past that ceiling.
Composer 2.5 is built on Moonshot AI's open-source Kimi K2.5 checkpoint, then trained further using targeted reinforcement learning with textual feedback, 25x more synthetic tasks than its predecessor, and novel infrastructure techniques developed in-house. Cursor did not need permission from any foundation model lab to ship this. They are not paying per-token API fees to a supplier for their core product. They own the model.
This matters beyond Cursor for one specific reason: the economics are now reproducible. Open-source frontier checkpoints are getting better fast. Compute partnerships (like the xAI/Colossus 2 arrangement) are available to well-capitalized startups, not just trillion-dollar labs. The custom RL techniques Cursor is publishing are accessible to any serious ML team. The barrier to running this playbook is falling.
What Cursor is demonstrating is that an application company, if it moves fast enough and builds the right training infrastructure, can decouple from the foundation model supply chain entirely.
The take
The most successful AI companies over the next three years will not be the ones that built the best wrapper. They will be the ones that used the wrapper phase to buy time, build distribution, collect proprietary data, and then climb the stack.
Cursor had one critical advantage that made this possible: it sat directly in the developer workflow. Every coding session was a training signal. Every accepted completion, every rejected suggestion, every rerun was data that no foundation model lab had in the same form. That proprietary behavioral data, combined with open-source base models and custom RL, is what produced a model competitive with Opus 4.7 at one-sixth the input token price.
The uncomfortable question for every AI product company right now is not "which model should we use?" It is "are we collecting the data and building the infrastructure to eventually not need anyone else's model?"
Most are not. And that gap will widen.
The number
$0.50/M — The input token price for Composer 2.5. Anthropic's Opus 4.7 and GPT-5.5 sit in a range that makes this look like a different category entirely. For a model that nearly matches them on SWE-Bench Multilingual, the pricing is not a discount strategy. It is a signal that vertical integration in AI has a real cost advantage, and Cursor is already there.