I asked Claude which database to use 612 times, then had Claude Code and Cursor build the same apps. In 4 of 7 jobs they built something else.
retained text
I wanted to know if what Claude recommends in chat matches what coding agents actually do when you let them choose. **Setup** * Chat: 612 answers from Claude Sonnet 5 (API) to questions like “What database should I use for a new web app with users and their data?” * Builds: 87 runs of Claude Code (Opus 5.5) and Cursor. Each got a real build task for the same use case in a fresh, empty project, with “choose whatever database you think is best”. * Oct 23 – Nov 3, 2026 **Results** (chat’s top pick → what the agents built) |Job|Chat recommends|Agents built| |:-|:-|:-| |New web app|Postgres 92%|SQLite 93% | |Python API|Postgres 92%|SQLite 58%, Postgres 33% | |Full-text search|Postgres 54%, Elasticsearch 35%|SQLite 92%, Elasticsearch 0% | |Analytics events|Postgres 61%, ClickHouse 38%|ClickHouse 83%| |Serverless on Vercel|Neon 33%|Neon 100% | |Self-hosted|Postgres 96%|Postgres 100%| |RAG / vectors|Postgres 73%|Postgres 100%| **My read:** chat gives the “textbook” answer. Agents pick whatever gets the app running with zero setup. No account, no connection string. For RAG, Pinecone came up in chat answers a lot but was recommended in only 1% of them, and the agents never used a dedicated vector DB. **Limitations** (please poke holes) * Small samples: 12–15 builds per job. * Chat and the agents were different models, on purpose: chat is what people ask, agents are what people build with. * The chat question and the build task describe the same use case but aren’t worded the same. * For a prototype, SQLite is arguably the right call. I’m not saying it’s wrong, just that it isn’t what chat tells you. Has anyone seen the same gap in other categories (auth, email, payments)?
document fetched 2026-11-04 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xmxjiy/i_asked_claude_which_database_to_use_612_times