Frontier labs verticalize as open source models revive harness value

Sahil Jain observes that frontier model labs such as OpenAI, Anthropic, and Google are unlikely to open-source their core models. Instead, these companies are pursuing vertical integration to own the end-to-end stack, including harnesses and applications. OpenAI has deployed engineers into banks and partnered with Base 10 to offer open models within its ecosystem, aiming to capture enterprise budget spend. In contrast, labs like NVIDIA are open-sourcing models like Nemotron to drive hardware adoption across a diversified customer base.

Jain argues that while model intelligence improvements reduce the need for complex harnesses to compensate for deficiencies, the rise of open-source models is reviving the value of harnesses for cost-efficient routing. He notes that model release cycles have accelerated, with the most recent state-of-the-art jump occurring in 11 days, compared to previous intervals of months. He also highlights that Anthropic's Claude Code performs better with Anthropic models than with other harnesses, suggesting co-evolution between models and their native tools. Jain posits that the future of non-model companies lies in building tools that work across diverse models without overfitting to a single provider, leveraging the cost efficiency of open-source models for production workloads. He summarized the dynamic by stating, "The model itself is eating the harness."