Open models beat frontier models when you customize them. Here's how.

The panel argues that open models have reached a critical inflection point: they now match or exceed closed frontier models when optimized for specific harnesses. The key insight is that generalized chatbot intelligence is overkill for 90% of tasks. Teams should focus on post-training open models on their exact use-case environment rather than paying premium

API costs for broad capability they don't need. The real leverage comes from owning the entire stack—data, model weights, and inference traces. Closed APIs obscure both cost and data usage. Open models let teams build a data flywheel: each production trace improves the specialized model. Companies like RAM showed they could surpass Opus in finance

by post-training an open model in two weeks at a fraction of Haiku's cost. The panel is skeptical about trust arguments used against open models. Closed APIs are less trustworthy because they hide training data and can deprecate models arbitrarily. Open models allow validation of weights, data, and behavior. The geopolitical concern about Chinese models

ironically reversed when Anthropic restricted access, driving enterprises back to open alternatives. Tradeoffs remain: frontier models still lead on multi-domain benchmarks. Not every task needs customization, and the engineering effort to set up RL environments is non-trivial. But the panel predicts that within 12 months, most daily AI tasks will run on-device, and computers will