AI models and tools
Models change every few months. Your job isn't to pick the best model once. It is to make switching cheap and to decide based on your own evidence.
| Category | Options (late 2026) |
|---|---|
| Primary coding agent | Claude Code with Claude Opus or Fable |
| Everyday and subagent work | Claude Sonnet |
| Fast, cheap tasks | Claude Haiku (summaries, classification, memory) |
| Second opinions | OpenAI Codex, Cursor, Gemini |
| Images and multimodal | Claude, Gemini, OpenAI image models |
The same model in different tools produces different quality, because system prompts and tooling matter. Use more than one agent to cross-check important work.
Staff decisions here:
- An approved list. Models and tools with reviewed data-handling terms. Engineers choose within it
- Your own evals. Benchmarks don't know your codebase. Run a sample of real tasks before switching
- Model IDs in config. Not scattered through code, so an upgrade is one change
- An exit path. Don't lock the org into one vendor's proprietary features without one