DefaultRank report · 2026-10-01
AI recommends one stack and builds another
We asked 6 AI models (claude-sonnet-5.5, gpt-6.1-sol, gemini-3.8-flash, grok-4.7, deepseek-v4.1-flash, qwen3.8-max-prime) which auth, database and email tools to use, and then asked them to build apps that need them. 749 answers later, the tools they recommend and the tools they actually put in code often don't match.
Recommended is not what gets built
Asked which auth provider to use, models pick Clerk first 62% of the time. Asked to build an app with sign-up, they put Clerk in 1% of apps.
AI-built apps mostly skip auth vendors
37% of AI-built apps hand-roll auth with password hashing and JWTs, and 35% use Auth.js. Hosted auth providers together appear in 9%.
Models build with what they learned
Auth.js is now maintained by the Better Auth team, which points new projects to Better Auth. In AI-built apps, Auth.js appears in 35% and Better Auth in 0%.
SQLite is the default prototype database
Recommendations favor PostgreSQL (74% of top picks), but 28% of AI-built apps use SQLite. Hosted database vendors are named in 37% of them.
Email is the exception
Resend is both the top recommendation (61%) and the most-used email API in AI-built apps (55%). The main alternative in code is plain SMTP via Nodemailer (32%).
Authentication
Top recommendation: Clerk (62%). Most used in AI-built apps: Custom (DIY) (37%).
Each model's favorite
| Model | Most recommended | Most used in its apps |
|---|---|---|
| claude-sonnet-5.5 | Clerk 50% | Auth.js (NextAuth) 47% |
| gpt-6.1-sol | Clerk 77% | Custom (DIY) 40% |
| gemini-3.8-flash | Clerk 60% | Custom (DIY) 30% |
| grok-4.7 | Clerk 70% | Custom (DIY) 47% |
| deepseek-v4.1-flash | Clerk 77% | Auth.js (NextAuth) 55% |
| qwen3.8-max-prime | Clerk 37% | Custom (DIY) 47% |
Database
Top recommendation: Supabase (54%). Most used in AI-built apps: Neon (21%).
Vendor view: an answer like “PostgreSQL, hosted on Neon” counts for Neon, and generic engines are left out. By engine, PostgreSQL is the top pick in 74% of recommendations; in AI-built apps PostgreSQL appears in 58% and SQLite in 28%.
Each model's favorite
| Model | Most recommended | Most used in its apps |
|---|---|---|
| claude-sonnet-5.5 | Supabase 67% | Neon 33% |
| gpt-6.1-sol | Supabase 50% | Firebase (Firestore) 17% |
| gemini-3.8-flash | Supabase 77% | Supabase 47% |
| grok-4.7 | Neon 50% | Neon 10% |
| deepseek-v4.1-flash | Neon 50% | Neon 45% |
| qwen3.8-max-prime | Supabase 53% | Neon 27% |
Top recommendation: Resend (61%). Most used in AI-built apps: Resend (55%).
Each model's favorite
| Model | Most recommended | Most used in its apps |
|---|---|---|
| claude-sonnet-5.5 | Resend 50% | Nodemailer (SMTP) 50% |
| gpt-6.1-sol | Postmark 53% | Resend 80% |
| gemini-3.8-flash | Resend 87% | Resend 90% |
| grok-4.7 | Resend 80% | Nodemailer (SMTP) 50% |
| deepseek-v4.1-flash | Resend 67% | Resend 75% |
| qwen3.8-max-prime | Postmark 63% | Nodemailer (SMTP) 70% |
AI writes code for last year's stack
Every one of these AI-built apps shipped at least one dependency a major version behind npm's latest. See which packages and which models, or check your own package.json.
Rarely surfaced
These tools were named in fewer than 2% of recommendations in their category. Our prompts are generic, so tools built for a specific niche are at a disadvantage here, and this is not a judgment of quality.
Appwrite Auth, Buttondown, Convex Auth, Descope, Frontegg, FusionAuth, Hanko, InstantDB, Kit (ConvertKit), Logto, Loops, Mailchimp, MailerSend, Mailjet, Mailtrap, Nhost, Nile, Passport.js, Prisma Postgres, PropelAuth, SingleStore, SparkPost, Stack Auth, SuperTokens, TiDB, Xata, ZeptoMail, beehiiv
How to read this
- Recommended first is the first tool a model names when asked what to use (6 wordings per category covering different developer contexts, each asked 5 times per model).
- Used in AI-built apps is what ends up in package.json, imports and config when a model is asked to build an app (7 app prompts, 5 times each per model, low reasoning effort).
- These are direct API calls with no system prompt. Coding agents add their own instructions, tools and docs, and may behave differently.
- Per-model figures rest on about 30 answers each, so treat differences of less than about 20 points between models as noise. Category totals are within about ±7 points.
- Results describe these model versions on 2026-10-01. Models change, so we re-measure regularly.
Full prompts, models and detection rules: methodology. Aggregated data: data.json. DefaultRank is independent and not affiliated with any tool vendor or model provider. No vendor paid for or reviewed these results.
Is your tool in one of these categories?
See instantly how AI models recommend and use your tool, then get the full breakdown by model and developer context.