BlogApplied AI

The AI Startups Use in 2026 (Y Combinator Data) and What Nobody Tells You

More than half of Y Combinator's latest batch uses Claude. The data, the sources, and what nobody tells you about AI-written code.

V

Vectorial Team

Published Nov 20, 2026

The AI Startups Use in 2026 (Y Combinator Data) and What Nobody Tells You

In short

In Y Combinator's Winter 2026 batch, more than 52% of startups use Anthropic's Claude models, up from about 25% in early 2025. A Supabase survey of about 2,000 founders shows 59% pay for Claude and 39% for ChatGPT. But 45% of AI-generated code fails security tests, so human review is still key.

A few weeks ago we told you which technologies Y Combinator startups use to build their websites. Now we're going a step further: which artificial intelligence they use to work. And there's a surprise.

Claude overtook ChatGPT among startups

On Y Combinator's Lightcone podcast (December 2025), YC partners said more than 52% of startups in the Winter 2026 batch use Anthropic's Claude models. In early 2025 it was about one in four.

A Supabase survey of about 2,000 startup founders (not only from YC) shows the same trend:

Paid subscriptionsBefore2026
Claude28%59%
ChatGPT / OpenAI57%39%
Gemini—27%

For coding tools, the same survey shows Claude Code at 63% usage, Visual Studio Code at 44%, and Cursor at 31%.

AI already writes most of the code

Y Combinator reported that 25% of startups in its Winter 2025 batch had codebases that were 95% AI-generated. For a traditional company, that means something important: what used to require a large team of developers can now cost far less.

What nobody tells you

According to Veracode's 2025 report, which tested more than 100 AI models on 80 coding tasks, 45% of AI-generated code fails security tests. AI writes fast, but it doesn't always write well. Without someone reviewing it, code can carry vulnerabilities nobody notices until it's too late.

How we use it at Vectorial

  • AI writes: Claude Code builds most of the code.
  • A person reviews: security, business logic, and design go through human review.
  • It's tested before release: nothing goes to production without verification.

AI doesn't replace judgment. It multiplies it.

Sources

  • Y Combinator, Lightcone podcast (December 2025), cited by Inc.
  • Y Combinator, Winter 2025 batch, cited by Cloud Security Alliance (April 2026).
  • Supabase, State of Startups 2026 (about 2,000 founders; Supabase is an industry vendor).
  • Veracode, 2025 GenAI Code Security Report.

Frequently asked questions

Related service: View custom software development