Lindy Switched 100% From Claude to DeepSeek — A Real Migration Cost Breakdown

The Migration That Actually Happened

On June 27, 2026, Lindy CEO Flo Crivello told reporters his San Francisco AI startup had moved 100% of its inference traffic from Anthropic's Claude to DeepSeek earlier in the month. His framing was blunt: the Claude bill had exceeded total employee salary spend, and the company expected DeepSeek to "save several million dollars over the coming months."

Lindy is not a tiny shop. They run automation agents at production scale across thousands of customers. A full provider switch at that scale is a serious vote of no confidence in frontier pricing — and it is rare enough that the math behind it deserves a look.

The Pricing Gap as of June 2026

Current published rates for the two most likely models involved:

DeepSeek V4-Pro is roughly 11x cheaper on input and 14x cheaper on output than Claude Sonnet 4.6. On a Lindy-scale workload of, say, 5B input tokens and 800M output tokens per month, the gap is real money: $27,000/month on DeepSeek V4-Pro vs $360,000/month on Claude Sonnet 4.6.

What Has to Be True for the Switch to Work

A 13x cost reduction looks irresistible on a spreadsheet, but the switch only nets out positive when three conditions hold.

Condition 1: Quality delta is within margin. On real Lindy-style automation tasks (form filling, summarization, light reasoning), DeepSeek V4-Pro benchmarks within 3-7 points of Sonnet 4.6 on SWE-Bench Verified and TerminalBench. That gap matters less than the cost ratio.

Condition 2: Prompt rewrites are bounded. Claude and DeepSeek diverge on system-prompt grammar (Anthropic's XML tag style vs DeepSeek's plain-text preference) and on tool-call schemas. Lindy almost certainly has 100+ prompts in production; rewriting them is a one-time cost that needs to be amortized.

Condition 3: Failover and SLA are in place. DeepSeek had a multi-day outage in Q1 2026. Switching 100% of traffic to a single provider in China without a fallback path is operational risk you might not want. Lindy's announcement does not detail their failover plan but anyone copying the move should plan for at least a 5-10% reserve budget on a Western provider.

The Hidden Switching Cost Most Teams Underestimate

Vendor migration is not just prompt rewrites. There are four cost lines that surface only after you commit.

• Eval rebuild: your golden eval set was probably tuned to Claude behavior; it needs to be re-baselined.
• Caching contract change: Claude has predictable prompt caching with a 1-hour beta TTL. DeepSeek uses prefix caching with different invalidation rules.
• Tool-call format: JSON schema parity is close but not identical. Edge cases break silently in production.
• Compliance review: for some enterprise customers, China-hosted inference is a hard no. Lindy did not break this out but it likely cost them some accounts.

A Decision Framework for Your Team

Use this to decide whether a Lindy-style switch makes sense:

Step 1: Measure annual run-rate spend on your current frontier model. If under $50K/year, the switching cost will likely exceed the savings.

Step 2: Run a 1-week shadow eval — send 5% of production traffic to DeepSeek in parallel, compare outputs on your real eval set. If quality delta is under 5%, the switch is viable on quality grounds.

Step 3: Estimate switching cost. Rule of thumb: 80 engineering hours per 50 prompts, plus 40 hours for eval rebuild, plus 20 hours for tool-call schema adaptation. At a $120/hour blended rate, that is roughly $20K of one-time spend.

Step 4: Compute payback period. If annual savings ÷ switching cost is under 6 months, do it. Lindy's payback is under one month given the scale.

The Takeaway

Lindy's switch is not a referendum on Claude's quality — Claude Opus 4.8 still leads frontier coding benchmarks. It is a referendum on pricing. When the cheapest credible alternative is 11-13x less expensive and within 5 points on the tasks that matter to your workload, the math becomes unignorable. Expect more public migrations through the rest of 2026. The teams that move first will lock in margin advantages that compound through the year.