The bargain
The shape repeats everywhere. A user request fires one or more API calls to a model provider; the response is wrapped in a UI and an integration; a bill arrives monthly. That bill is the dominant cost of goods sold — often most of revenue early on. The model is a commodity input. The product is everything built around it: retrieval over private data, integration into a system of record, evals and guardrails, and the workflow logic that makes the output something a job depends on.
The map below is a curated, fact-checked roster of app-layer companies (mid-2026). Filter it by vertical and by model strategy — and watch how few actually train a base model. Click any company to see what it built and how it keeps the token bill from eating the margin.
Almost none train a base model. Most route or orchestrate someone else’s; a few fine-tune on rented weights. Click a band to filter the map below.
Cursor / Anysphere
Agentic coding IDE (VS Code fork) — multi-file edits, test loops, background agents.
Claude + GPT for hard reasoning; its own Composer (built on Moonshot’s open-weights Kimi K2.5 + Cursor RL) for fast edits.
The agent loop, the editor UX, and a model it built on open weights to survive its own inference bill.
The playbook the roster reveals
The COGS trap. If the product is one model call with thin wrapping, gross margin is capped at price minus token cost — and the provider can compress that to zero, because they buy the model at cost while you pay retail. Reselling raw tokens is renting a business from a competitor.
Margin lives outside the model. Proprietary or licensed data, integration into a system of record, evals and guardrails, and workflow logic. None of it gets cheaper when the model does — it gets more valuable.
Decouple price from tokens. Per-seat or per-outcome pricing turns a falling token cost into expanding margin instead of falling revenue. Same price to the customer, lower cost to serve — the support agents billing per resolution are the cleanest example.
Routing is leverage.Send easy, high-volume work to the cheapest adequate model and reserve frontier models for the hard tenth, so no single provider can hold you hostage on price or capability. (It's the exact problem the rest of this site exists to measure.)
Training your own model is an escape hatch, not a default.Worth it only when volume is enormous, the task is narrow enough to train cheaply, and the frontier isn't already beating your custom model. Cursor built one (on open weights) to survive its bill; Harvey built one and threw it away when frontier models lapped it. Most should route.
Platform risk — what actually kills these companies
The structural danger is that the model provider ships your product natively. When a major lab released a general browser-automation agent, startups whose whole value was browser automation lost their reason to exist. The threat is the unlit square on the upstream vendor's roadmap, not the competitor across the street.
What survives is the layer the provider doesn't want to build: regulated-industry compliance, proprietary data, integration into systems of record, and the trust that comes with them. The durable position is vertical depth the model owner has no incentive to chase — not a thinner, prettier version of their chat box. On the map above, those are the companies whose moat line never once mentions the model.