Vertical AI Models and Post-Training on Usage Data

Vertical AI Models and Post-Training on Usage Data

Definition

The emerging pattern where product companies build domain-specific AI models by post-training open-source base models on proprietary usage data ("last-mile" interactions), achieving performance that matches or exceeds general frontier models at lower cost. Represents a shift from relying on API calls to frontier labs toward full-stack ownership (app + AI + model).

Key Points

  • Cursor's Composer 2 built on Kimi K2.5 with RL post-training on coding interaction data; beats Opus 4.6 and matches GPT-5.4 on coding benchmarks at lower cost; 75% of compute from Cursor's proprietary training (ai daily brief anthropic mythos vertical models)
  • Intercom's Fin Apex (dedicated customer service model) beats GPT-5.4 and Opus 4.5 on resolution rates with 65% fewer hallucinations and dramatically lower cost (ai daily brief anthropic mythos vertical models)
  • Decagon runs 80%+ of model traffic on in-house models using a network of specialized models for different customer service stages (detection, orchestration, response, evaluation) optimized independently (ai daily brief anthropic mythos vertical models)
  • Multiple product companies (Pinterest, Airbnb, Notion, Cursor, Intercom) are finding it better/cheaper/faster to post-train open models in-house rather than rely on API calls (ai daily brief anthropic mythos vertical models)
  • This pattern is distinct from early domain-specific pre-training (Bloomberg GPT) which was "smoked by general models" — the key difference is post-training on proprietary usage data (real user interactions) rather than just domain text (ai daily brief anthropic mythos vertical models)
  • Paul Adams (Intercom CPO) framing: vertical models can and will outperform general models; durable differentiation moves down the stack to the model layer; successful companies must be full-stack across app, AI, and model layers (ai daily brief anthropic mythos vertical models)
  • Eoghan McCabe references Andrej Karpathy's "speciation" prediction: smaller models with cognitive core, optimized for specific tasks (ai daily brief anthropic mythos vertical models)
  • Richard Sutton's "Bitter Lesson Revisited" on Dwarkesh podcast may have predicted this: systems trained from experience supersede those built on human knowledge — vertical models are post-trained from experience (usage data) (ai daily brief anthropic mythos vertical models)
  • "API tax starting to look like the cloud markup of 10 years ago" — observation that per-token pricing to frontier labs becomes uneconomical compared to in-house training (ai daily brief anthropic mythos vertical models)
  • Cursor's Composer 2 pricing at $0.50/M input tokens and $2.50/M output tokens (March 2026) is "substantially below frontier model pricing from Anthropic and OpenAI" — token economics drive tool selection more than feature comparisons for teams running dozens of parallel agents (cursor 3 agent management console)
  • Cursor responded to Claude Code overtaking its revenue ($2.5B vs. $2B run rates) with a six-week product offensive including Composer 2 (in-house vertical model), demonstrating how competitive pressure accelerates vertical model adoption (cursor 3 agent management console)

Open Questions

  • Will every product company with sufficient usage data train vertical models, or will this remain limited to well-funded companies with ML expertise?
  • What is the minimum dataset size / interaction volume needed for vertical model training to outperform general models?
  • How will frontier labs (OpenAI, Anthropic, GDM) respond to customer attrition from vertical model adoption?
  • Will vertical model tooling (training pipelines, eval harnesses, deployment infra) become its own product category?
  • Does this shift reduce or increase demand for GPUs? (Less inference API calls, but more training runs)
  • Can vertical models maintain quality as base models improve, or will they need continuous re-training?
  • What happens to companies built entirely on frontier API access if the economic moat shifts to vertical models?
  • For Microsoft/GitHub: should Azure AI provide "vertical model training as a service" to capture companies making this shift?

Related Concepts