As API fees from OpenAI and Anthropic continue to climb, a growing number of AI application startups are embracing open-source weight models to regain cost control. Legal AI unicorn Harvey, whose gross margins plummeted from approximately 50% at the beginning of the year to -50% in June, was prompted to release its own model in August, based on Moonshot AI's Kimi K3. This move, costing a fraction of Anthropic's top-tier product, has already returned Harvey to positive gross margins. Healthcare technology company Abridge, AI customer service company Decagon, fintech companies Ramp and Rogo, among others, have also announced plans for self-developed or customized models, with some already switching 80% of their traffic to their proprietary models. This trend puts direct pressure on OpenAI and Anthropic, which are preparing for IPOs, as the outflow of application-layer customers could erode their core revenue streams. Leading investment firms such as Sequoia Capital and General Catalyst are also driving this shift, with some investors even stating that companies not considering building their own models may struggle to secure funding. Furthermore, OpenAI and Anthropic's recent implementation of additional usage fees for enterprise users and their aggressive entry into core startup sectors have intensified competition. Although transitioning to open-source models presents challenges such as talent, data, and infrastructure costs, the industry generally believes that a hybrid strategy of using both open-source and closed-source models will be the optimal solution for most startups currently.