Claude writes better Spring Boot configs than your senior dev

I've benched all four majors against a nasty legacy Java codebase — 2,000 lines of spaghetti services, circular dependencies, and a UserService that somehow manages payment processing. Here's what actually happened.

Architecture decisions: Claude 3.5 Sonnet > GPT-4o > Gemini 1.5 Pro > DeepSeek-V3

• Context retention on multi-file refactors: Sonnet held 15-file context for 45 minutes. GPT-4o started forgetting package structures after file 8. Gemini hallucinated a Repository interface that didn't exist. DeepSeek gave up and suggested rewriting in Go.
• Spring Boot specifics: Sonnet correctly identified @EnableJpaRepositories vs @EntityScan confusion in 3/3 tests. GPT-4o conflated them twice. Gemini suggested @EnableAutoConfiguration as a fix for everything. DeepSeek didn't know @Transactional propagation types.
• PostgreSQL schema design: Sonnet produced a normalized schema with proper indexes and foreign keys on first try. GPT-4o forgot ON DELETE CASCADE on a join table. Gemini suggested JSONB for everything including primary keys. DeepSeek generated MySQL syntax.

# Prompt: "Add optimistic locking to this JPA entity with version field"

Sonnet: Added @Version, explained OptimisticLockException handling in service layer, warned about detached entity merges.

GPT-4o: Added @Version, missed the service-layer retry logic.

Gemini: Added @Version on a String field. Compiled. Failed at runtime.

DeepSeek: Added @Version, then rewrote the entire entity in Kotlin unprompted.

Where you still need a human (or at least a very careful prompt engineer)

• Redis cache invalidation strategies: Every model suggests @CacheEvict(allEntries = true) like it's free. It's not. Sonnet at least warns you when you ask for it on a 500k-entry cache.
• Database migration ordering: Flyway/Liquibase scripts generated by LLMs assume clean slate. They don't handle "this column existed in prod for 3 years with nullable=true now you're making it not-null."
• Angular change detection zones: All four models default to ChangeDetectionStrategy.Default and wonder why your 500-row table lags. Sonnet is the only one that proactively suggests OnPush + trackBy without being asked.

1. Let Sonnet design the Spring module boundaries — feed it your domain description, ask for a package diagram with bounded contexts. It'll give you something defensible.
2. Write your own Flyway migrations — LLMs don't know your production data.
3. Use Sonnet for Angular component scaffolding + OnPush patterns — it genuinely understands the zone.js trap.
4. Redis? Write the cache keys yourself. Naming conventions (user:123:profile:v2) are tribal knowledge no model has.
5. Hire a senior for 2 hours — not for code, for a 30-minute architecture review and a 90-minute "here's where you'll regret this in 6 months" session. Worth every penny.

The app is buildable. The stack is fine. But don't pretend Claude replaces architectural judgment — it just makes the iteration loop tighter when you have judgment.