Hopp til hovedinnhold
AIKI

Lindy dropped Claude for DeepSeek V4: A lesson for SMBs

||4 min lesing

Key takeaways

  • Lindy moved 100 percent of user-facing production from Anthropic Claude to DeepSeek V4 hosted in the US
  • AI inference was the single largest cost, larger than payroll. CEO Flo Crivello called it 'a matter of survival'
  • Vercel AI Gateway: DeepSeek's token share jumped from 1 to 17 percent in one month, while its share of total cost stayed near 1 percent

Lindy dropped Claude for DeepSeek V4: A lesson for SMBs

A 25-person AI agent startup in San Francisco has moved 100 percent of its user-facing production traffic from Anthropic Claude to DeepSeek V4, and according to the founder, the company is saving millions of dollars. The case is about price pressure, not ideology, and it matters for SMBs evaluating AI agents in 2026.

What actually happened at Lindy, and why did they pick DeepSeek V4?

Lindy is a no-code platform for AI agents that automates email triage, meeting scheduling, and CRM workflows. The company was founded in 2023 after a pivot from Teamflow. In June 2026, founder and CEO Flo Crivello (ex-Uber) announced on X that all production was moved from Claude to DeepSeek V4. Direct quote: "Saves us millions of $ and we're actually seeing an increase in performance on many core use cases. Transformative for the business."

What makes the case unusual is the scope. Crivello did not switch part of the traffic or run A/B tests. One hundred percent of what end users see now runs through DeepSeek V4. Internal employee use, where the flat-rate Claude Max subscription fits better, stays with Anthropic. The reasoning is consistent: AI inference was Lindy's largest single cost, in fact larger than payroll. Crivello called it "unsustainable" and "a matter of survival for the business".

How much did they save, and is it realistic for others?

Crivello himself states "millions of dollars" in savings, and CNBC reports the company will save millions "within months". There is no precise, verified dollar figure in open sources. What is known is that AI inference was Lindy's largest cost line, and that the cost curve, according to Crivello, "crashed to the ground" after the switch.

For SMBs, the right read here is a cost lens, not a copy-paste action. Lindy is a special case: an AI agent platform with extremely high inference traffic per employee, where every customer interaction pushes tokens through the model. A 15-person law firm that uses AI to summarize case documents once a week has a completely different cost profile. The lesson from Lindy is not about copying the saving. It is about understanding the mechanism: when AI becomes an ongoing operating cost, model choice becomes a strategic decision.

Why DeepSeek V4, and not something else?

Crivello evaluated open source models for 6 to 9 months before landing on DeepSeek V4, and spent another two months on DeepSeek-specific testing. Self-hosting was considered but rejected as a "massive distraction" for a 25-person team. The solution was US hosting: DeepSeek V4 runs as an open-weight model through a US vendor on US soil. Data never leaves the US.

The market alternatives split into two groups: premium closed models from OpenAI and Anthropic, and dramatically cheaper open-weight alternatives. Vercel's AI Gateway data from May 2026 shows that DeepSeek's share of token volume jumped from under 1 percent to 17 percent in one month, while its share of total cost stayed near 1 percent. That is a direct measurement of the price gap. For an SMB that wants to dig into pricing and model comparison, we have previously written about DeepSeek V4 and AI pricing and about Chinese language models in the portfolio. Crivello's comment captures the logic: he would switch back to Anthropic if they cut price meaningfully, but for now he has alternatives.

What does this mean for SMBs?

For an SMB with 7 to 100 employees, the question is not whether to switch from Claude to DeepSeek tomorrow. The question is whether you have a deliberate model portfolio, or whether you have locked all AI use to one premium vendor without knowing what it costs you over time. For a deeper walkthrough of models available in 2026 and how to choose an AI vendor, we have separate overviews. Three observations from the Lindy case that hit SMBs directly:

First: AI cost scales differently from software cost. When AI agents take over workflows, token consumption grows with usage. It is an operating cost, not a license cost. Second: the vendor market is moving. OpenAI is delaying its IPO because it refuses to go below a trillion-dollar valuation, while Snowflake CEO Sridhar Ramaswamy recently tested GLM-5.2 from Zhipu against Claude Opus 4.7 on 103 coding tasks. GLM solved 66 percent versus Opus's 67 percent across three attempts, at a fraction of the price. Third: many companies sit with a "just use Claude" default without having done a deliberate evaluation. That is no longer the safe answer. It is just an unevaluated position. A review of what AI agents actually cost in Norway in 2026 gives a realistic starting point.

The solution Lindy chose is geopolitically neutralized without being politicized. Open-weight means the model weights are openly available, but the inference runs at a US cloud provider on US soil. That means Norwegian customer data is processed under US jurisdiction, not Chinese. For SMBs without personal data beyond standard work data, this is a very different risk profile than sending data to a closed Chinese platform.

What remains is the GDPR and privacy assessment, which is independent of model choice. For SMBs handling health, finance, or other sensitive personal data, hosting choice, data processing agreement, and logging are still critical. We have written about AI agents, local hosting, and governance for those who want to go deeper on architecture choices. The point is that the Lindy case shows "open model" and "safe hosting" are not opposites. They are two independent choices.

How can AIKI clients put this to work?

At AIKI we start with a cost analysis of existing AI use. Many clients discover they are paying for Claude or GPT-4o on tasks that a cheaper model handles just as well: email categorization, first drafts of customer letters, meeting notes, simple code help. Then we set up a model portfolio where premium is used only where it actually pays off, and cheaper models take the volume.

For clients with high inference volume, building their own AI agents in production, we look at open-weight alternatives like DeepSeek V4 or GLM-5.2, always hosted in the EU or US and never through closed Chinese platforms. For clients with low volume and high consistency requirements, we often keep the premium vendor, but structure usage so that volume tasks do not pull unnecessary cost. The result is rarely one model for everything. It is a deliberate mix.

For a deeper introduction to AI agents in a Norwegian context and how they differ from general AI use, we have a separate overview.

Want to know what your current AI bill actually consists of, and where it can be cut without losing quality? Book a call with AIKI, and we will go through your actual usage and invoice.

Sources

Del:LinkedInXFacebook

Relaterte innlegg

Nyheter · 16. juni 2026Anthropic pauses Agent SDK change: 3 steps for SMBsNyheter · 15. juni 2026Anthropic Agent SDK June 15: 3 Steps for Norwegian SMBsNyheter · 9. juni 2026Claude Fable 5: Anthropic's new top model costs 2x