Opus 4.7 Is a Backend Model
Ask any developer what Claude is good at and you’ll hear “frontend, React, UI.” The data says otherwise.
What Opus 4.7 actually builds
When we run LLM-assisted classification on a random sample of 413 Claude Opus 4.7 repos, this is the primary-domain distribution:
| Domain | Count | Share |
|---|---|---|
| Web backend | 189 | 46% |
| Frontend | 119 | 29% |
| CLI tool | 51 | 12% |
| Data pipeline | 23 | 6% |
| Game | 19 | 5% |
| ML training | 17 | 4% |
| Infrastructure | 8 | 2% |
| MCP server | 2 | <1% |
Web backend alone is 46% — more than frontend + CLI combined.
Why the perception gap?
Three reasons people think of Claude as a “frontend model”:
- Demos are visual. Frontend apps look impressive in a 30-second screencast. Backend services don’t.
- Claude’s strengths show clearest in UI work. Component structure, Tailwind classes, accessibility — these are easy wins.
- Survivorship bias. The repos that go viral on Twitter are the beautiful frontends. The 189 backend repos in our sample mostly don’t go viral; they just quietly ship.
What kind of backend?
Diving into the 189 web-backend repos:
- RESTful API endpoints is the most common pattern — appears in 44 of 189 (23%)
- JWT authentication appears in 55 repos (29%)
- MVC architecture — 36 repos
- Dockerized deployment — 29 repos
- ORM-based database access — 12 repos
Framework split within backend repos:
- FastAPI — 225 mentions (leader)
- Express — 130
- Flask — 47
- SQLAlchemy — 90 (ORM layer)
- Prisma — 71
- Drizzle — 50
So when Opus 4.7 writes a backend, the default is FastAPI (Python) or Express/Next.js API routes (TypeScript), with Prisma/Drizzle/SQLAlchemy for persistence and JWT for auth.
CLI tools: the underrated category
12% is CLIs — a surprisingly high share. Names like:
screenase(Python CLI)chris2ao/unifi-mcp(Python MCP CLI — quality 84.5)astroicers/CyPulse(Python CLI, quality 82.2)
The highest-quality Opus 4.7 repos we’ve found are disproportionately Python CLIs, often MCP servers. This matches the hypothesis that small, well-scoped programs let the model show its best work.
Mobile and desktop are rare
Only 85 mobile apps (Expo, React Native, Flutter) and 30 desktop apps (Tauri). The model can do these — it just rarely does, at least in the publicly-committed corpus.
What this means
For anyone benchmarking Opus 4.7 or building on top of it:
- Don’t optimize only for frontend. Nearly half the real-world output is backend.
- FastAPI is the default Python framework. If your training data skews toward Flask, you’re out of distribution.
- CLIs are where the quality ceiling lives. The highest-graded Opus 4.7 repos are small Python CLIs. That’s a good benchmark target.
This post draws from 413 LLM-analyzed Opus 4.7 repos out of 9,281 indexed. Analysis continues; numbers update every 30 minutes.