This AI Agent Startup Ditched Anthropic for DeepSeek — and Says It’s Saving Millions

Why It Matters

The move highlights how soaring inference costs are forcing AI‑driven firms to prioritize cheaper, open‑weight models, reshaping vendor relationships and profit margins across the industry.

Key Takeaways

• Lindy switched from Anthropic to DeepSeek v4 for all traffic.
• Migration saved millions and improved performance on core use cases.
• Evaluation spanned 6‑9 months; rollout required 100x more work.
• DeepSeek’s cheap tokens split AI model market into premium and open‑weight.
• Lindy hosts DeepSeek via US‑based Atlas Cloud for data sovereignty.

Pulse Analysis

Rising inference costs have become the primary barrier to scaling AI services, prompting companies to scrutinize every token they consume. Lindy’s decision to replace Anthropic with DeepSeek v4 underscores a broader industry shift: enterprises are now willing to overhaul their model stacks to achieve multi‑million‑dollar savings. By moving to a model that delivers comparable results at a fraction of the price, Lindy not only improves its bottom line but also gains a performance edge on high‑frequency tasks such as email triage and meeting scheduling.

DeepSeek’s emergence signals a geopolitical and economic inflection point in the AI landscape. Developed by a Chinese research firm and optimized for Huawei’s CANN accelerator, the model delivers frontier‑class capabilities without relying on US‑based GPU infrastructure. This has accelerated the market bifurcation into ultra‑premium providers like OpenAI and Anthropic and a rapidly expanding tier of open‑weight alternatives that offer dramatically lower token costs. Vercel’s AI Gateway data, showing DeepSeek’s token volume jump to 17% while spending remains near 1%, illustrates how price‑sensitive workloads are gravitating toward these cheaper options.

For AI startups and venture‑backed platforms, Lindy’s experience serves as a cautionary tale and a roadmap. Companies with substantial token consumption must weigh the operational overhead of model migration against the long‑term financial upside. Hosting choices, such as using US‑based Atlas Cloud for DeepSeek, also address data‑sovereignty and compliance concerns that arise with Chinese‑origin models. As model pricing continues to evolve, flexibility and a willingness to experiment with open‑weight solutions will likely become a competitive necessity.