Lindy founder Flo Crivello announced on June 26 that the AI-agent platform has cut over 100% of its production traffic from Anthropic's Claude models to DeepSeek v4, hosted in the United States through inference provider Atlas Cloud. Crivello told CNBC the move saves Lindy 'millions of dollars' and called the prior trajectory of inference spend 'unsustainable' — AI bills had grown to exceed personnel costs at the 25-person startup. DeepSeek v4-Pro currently prices output tokens at about $3.48 per million; the company's Artificial Analysis Intelligence Index benchmark cost is roughly $1,071, versus $4,811 for Claude Opus 4.7.
Crivello's second claim matters as much as the cost figure: Lindy is seeing higher performance on many core agent workflows after the switch, not a regression. The migration was not instant — Crivello said it took six to nine months of evaluation, gradual rollout, and significant prompt re-engineering. That detail is the hidden tax of any model swap and the reason cheap weights do not, on their own, peel a customer off a frontier vendor. Lindy paid the migration cost because the post-switch math justified it.
The story is being read as the sharpest concrete signal yet of an inflection in the inference market. CNBC framed the broader move as 'from tokenmaxxing to efficiency' — buyers who in 2025 reached for whatever model was best are now optimizing dollars-per-task. Atlas Cloud hosting DeepSeek v4 on American soil also matters: it removes the national-security objection that has dogged direct China-API integrations and gives U.S. operators a cleaner path to use the cheapest competent weights available. Crivello said he would consider returning to Claude if prices come down, which is exactly the leverage the open-weight tier now has over the closed frontier.
A takeaway for learners: model selection is becoming an exercise in unit economics, not just leaderboards. If you are building anything with API calls in the critical path — agents, RAG pipelines, batch summarization — start tracking dollars-per-successful-task today, not just latency and quality. The teams that can swap providers without breaking their app will have meaningful pricing power. The teams that cannot will pay frontier rates whether or not they need them.