Skip to content
News Product launch

Lindy migrated 100% of its AI agent traffic from Claude to DeepSeek and cut inference costs by 90%

· by Pondero Newsdesk

The short version

The 25-person AI agent startup moved its managed-agent model traffic to DeepSeek v4 Flash on Atlas Cloud, citing inference costs that had surpassed personnel costs and called the switch a matter of survival.

Lindy migrated 100% of its AI agent traffic from Claude to DeepSeek and cut inference costs by 90%

AI agent platform Lindy moved all managed-agent model traffic from Anthropic's Claude to DeepSeek v4 Flash hosted on Atlas Cloud. The result: inference costs on the migrated routes fell roughly 90%, per a June 24 technical writeup by Lindy staff software engineer Bruno Skvorc. CEO Flo Crivello framed the decision bluntly to CNBC: "It's a matter of survival for the business."

What happened

Lindy is a 25-person startup that builds AI assistants to handle email, calendar, meetings, and follow-up tasks. The company had designed its pricing model around a bet that inference costs would fall fast enough to sustain a generalist AI product. By mid-2026, the bet was running behind: AI inference spend had exceeded Lindy's personnel costs, a threshold Crivello described as "unsustainable" per The Decoder's June 26 coverage.

The technical team ran offline evaluations across several candidate models, including GLM5.1, Kimi K2.6, and DeepSeek v4 Flash, plus testing across multiple inference providers. Lindy found that the same nominal model could score differently depending on the provider, which it attributes to quantization differences or provider-side inference stack variations. DeepSeek v4 Flash on Atlas Cloud won for the workloads Lindy cared about. One candidate, Kimi K2.5, passed offline evals but failed in a real-usage rollout when a user reported the assistant felt like "Lindy had had a brain surgery overnight."

After prompt-optimization work via the company's internal GEPA loop, Lindy rolled out DeepSeek incrementally, starting with internal employees and monitoring online evaluation signals and retention data for several weeks before ramping to 100% of the target traffic. Claude and Sonnet paths remain available when users explicitly select them or when a higher-intelligence route is needed.

Why it matters

The 90% cost reduction on migrated traffic is the headline number, but the more important detail for AI-tool operators is the migration process itself. Lindy's writeup is a rare public account of what a production-grade model swap actually requires: an offline eval suite, multi-provider testing, prompt re-optimization, a staged rollout, and weeks of retention monitoring before declaring success. Skipping any of those steps, as Lindy's Kimi K2.5 experience shows, produces a product that feels subtly broken even when benchmark scores look fine.

For operators running agents on Anthropic's Claude today, this is a concrete cost-pressure signal. Crivello said he would switch back if Anthropic cut prices, which suggests Lindy's move was economic, not a permanent architectural preference. The question is whether Anthropic responds with pricing changes. If it does not, other mid-sized agent platforms face the same arithmetic.

Lindy itself remains a worthwhile evaluation for teams building automated inbox, calendar, or meeting workflows. Try Lindy to see the product that emerged from this cost discipline exercise.

Context

CNBC placed Lindy inside a broader June 26 story about startups shifting from frontier models toward efficiency-first alternatives. Lindy's own post frames the dynamic from the application side: frontier models command a premium only where the quality gap is measurable. Below that threshold, cheaper alternatives absorb the traffic.

Per the Lindy blog post, "The pressure is on margins, not relevance." Labs still matter at the frontier; it is the tier just below the frontier that keeps commoditizing workload by workload.

What to watch next

Two signals to track. First, whether Anthropic adjusts enterprise pricing in the next 60 days in response to the public pressure this story represents. Second, whether similar migration announcements follow from other AI-native platforms of similar size. Lindy named n8n, Cursor, and Pipedream as natural comparisons in the market segment; public disclosure from any of them would confirm this as a trend rather than a company-specific outlier.

Sources