Lindy AI drops Claude for Deepseek, saves millions

Key Takeaways

• Lindy AI abandoned Claude entirely for Deepseek hosted on US infrastructure, saving millions in API costs
• AI spending at the 25-person startup had grown to exceed total personnel costs before the switch
• CEO Flo Crivello says he would return to Anthropic if they cut prices, framing it as a survival decision

Lindy AI, a 25-person startup building AI assistants, has completely replaced Anthropic's Claude with Deepseek. The switch saved the company millions, CEO Flo Crivello told CNBC, after AI costs had ballooned past what Lindy spent on salaries.

Crivello didn't mince words about the decision: "It's a matter of survival for the business." He added that he would switch back if Anthropic cut prices. That conditional loyalty underscores a brutal reality for frontier AI labs. Performance leadership means little if customers can't afford to use your models at scale.

Why AI costs hit a breaking point

Lindy builds AI-powered personal assistants that handle scheduling, email, and workflow automation. These tasks require heavy API usage. Each user interaction burns tokens, and agentic systems, where models call other models in loops, multiply that consumption fast.

OpenAI CEO Sam Altman recently acknowledged that AI costs have become a "huge issue" for companies deploying agentic workflows. When a single user session might involve dozens of chained API calls, the math gets ugly quickly.

For Lindy, the numbers became untenable. A startup with 25 employees spending more on AI inference than on people is a startup with a unit economics problem. Deepseek's pricing, roughly 90% cheaper than Claude for comparable tasks, made the decision straightforward.

Deepseek on US infrastructure

Crivello specified that Lindy runs Deepseek through a US company on US soil. This matters. Chinese AI models face scrutiny over data handling and potential regulatory risk. Hosting through a domestic provider sidesteps some of those concerns, even if the model weights originated from a Beijing lab.

The approach reflects a pragmatic middle ground. Companies get Deepseek's cost advantages without sending data to Chinese servers. Several US cloud providers now offer Deepseek inference as a managed service, making this path accessible to any startup willing to test it.

How Chinese models stack up

A recent analysis by Snowflake's CTO found that affordable Chinese models like GLM-5.2 don't quite match Claude's peak capabilities. But they're competitive. And for many production workloads, the price-performance ratio favors the cheaper option.

This is the uncomfortable truth for Anthropic and OpenAI. Not every task needs the smartest model. Most production AI systems spend tokens on routine completions where a model that's 85% as good at 10% of the cost wins decisively. The frontier matters for benchmarks. Price-performance matters for invoices.

Pressure on Anthropic's IPO timeline

Anthropic has been building toward an IPO, riding explosive revenue growth. But customer defections like Lindy's complicate the narrative. If high-volume users migrate to cheaper alternatives, growth projections get harder to defend.

OpenAI reportedly missed its optimal IPO window and postponed. Anthropic may face similar timing pressure. The longer pricing erosion continues, the harder it becomes to justify premium valuations based on API revenue alone.

Amazon has invested over $4 billion in Anthropic, with a significant stake tied to the company's success. That capital buys runway, but investors eventually want returns. A race to the bottom on inference pricing wasn't part of the pitch deck.

What this signals for AI builders

Lindy's move is a leading indicator, not an anomaly. Startups burning venture capital on AI inference will all face this calculation. When does model quality justify 10x the cost? For agentic systems running hundreds of calls per session, the answer increasingly looks like "rarely."

Crivello's willingness to return to Claude if prices drop suggests these decisions aren't permanent. Model providers retain leverage if they can close the cost gap. But that requires accepting lower margins or achieving dramatic efficiency gains.

Frequently Asked Questions

Why did Lindy AI switch from Claude to Deepseek?

Lindy's AI costs had grown to exceed total personnel costs, making the business unsustainable. Deepseek offered comparable capabilities at a fraction of Claude's price.

Is Deepseek as good as Claude for production use?

Analysis suggests Deepseek doesn't match Claude's peak performance, but it's competitive for many tasks and wins on price-performance ratio for routine workloads.

How is Lindy using Deepseek while avoiding Chinese data concerns?

Lindy runs Deepseek through a US company on US infrastructure, keeping data within domestic providers while using the Chinese-developed model weights.

Will Anthropic lower Claude's pricing?

Unknown. Lindy's CEO said he would switch back if Anthropic cuts prices, but the company faces pressure to maintain margins ahead of a potential IPO.

What does this mean for AI startup costs?

Agentic AI systems burn tokens rapidly, making inference costs a major expense. Startups may need to choose cheaper models or build hybrid architectures using frontier models selectively.

Source: The Decoder / Matthias Bastian