Jesse Zhang, the CEO of customer service AI startup Decagon, said on X that his company runs roughly 90 percent of its production workloads on open-source models rather than frontier systems from OpenAI or Anthropic. Zhang’s explanation cuts against the industry’s default assumption that bigger, more capable models win by default. The stakes are a live debate inside every enterprise AI buying committee: whether to standardize on a single frontier API or invest in smaller, purpose-built models once a use case is understood.
Zhang’s argument rests on a specific tradeoff. Small models that have been heavily fine-tuned for a narrow task, in Decagon’s case customer service agents, deliver lower latency and better task-specific accuracy than general-purpose frontier models. Customer service is a latency-sensitive category: a support agent that takes several seconds to respond breaks the experience regardless of how smart the underlying model is. Fine-tuning lets a small model compress broad capability into narrow competence, trading away flexibility for speed and reliability on one job.
The reason most companies still default to frontier models, per Zhang, is that most enterprise AI deployments are still early. When a team does not yet know the exact shape of its workflow, a frontier model’s breadth is valuable: it can handle edge cases, ambiguous inputs, and shifting requirements without retraining. Flexibility is worth paying a latency and cost premium for while the product is still being defined.
Zhang’s forecast is the more consequential part of the post. He expects that once workflows settle and mature, production traffic will shift away from closed frontier systems and toward specialized open-weight alternatives, following the path Decagon has already taken. That is a maturity curve, not a one-time architecture choice: flexibility matters most at the start of a deployment, and efficiency matters most once the task is understood well enough to fine-tune for it.
The framing does not include benchmark data or usage figures beyond the 90 percent workload split Zhang cited for his own company, so the claim should be read as one operator’s account of his own stack rather than an industry-wide measurement. Decagon has a business incentive to make this case: a company selling customer service agents benefits from a narrative where specialized tuning, its core competency, beats generic intelligence.
If Zhang’s pattern holds across other latency-sensitive categories such as voice agents, real-time fraud detection, and other narrow production workflows, the implication for frontier labs is direct. Revenue from frontier model APIs would concentrate more heavily in the early, exploratory phase of enterprise deployments, with the recurring, high-volume production traffic shifting to open-weight models that companies fine-tune and host themselves once the workflow is proven. That would not eliminate demand for frontier models, since new use cases keep entering the early stage, but it would cap how much of an enterprise’s steady-state inference spend a closed frontier API can expect to keep once a customer’s product ships.
Enterprise buyers evaluating a new AI workflow in the next ninety days should budget for a two-phase model strategy from the outset: a frontier model to prototype and validate the task, with a fine-tuning and evaluation plan already scoped for migrating that workload to a smaller open-source model once volume and requirements stabilize.
Decagon CEO Jesse Zhang, posted on X on July 6, 2026.
