Marketing pages are useless. We benchmarked these on our own workload: refactor a 4,000-line service, fix a bug from a stack trace, write tests for an unfamiliar module, explain a migration path.
Claude Opus 4: the best all-rounder. Fewer hallucinations. Best at refactoring legacy code with confidence.
GPT-4.1: the most predictable API. Structured outputs are the cleanest we've seen. Best when you need to trust the shape of the response.
Gemini 2.5 Pro: the context king. Feed it a whole codebase and ask 'where does this field get set'. It will answer.
Grok 4: the tool-calling champion. It's the strongest agentic loop we've benched. Worth a look if you're building autonomous agents.
The honest take: they're all frontier. The differentiator is the tooling around the model, not the model itself. Pick by ecosystem, not by leaderboard.