Choosing a Model
Pricing changes faster than most documentation gets updated. Use this page for decision logic, not static dollar figures.
What to Optimize For
Section titled “What to Optimize For”| If your main need is… | Optimize for… |
|---|---|
| Long-running agent loops | reasoning quality and tool use |
| Quick edits and completions | latency |
| Visual/UI implementation | multimodal strength |
| Sensitive code | local execution or trusted provider boundaries |
| Large investigations | long context plus strong context hygiene |
Workflow-Fit Rules
Section titled “Workflow-Fit Rules”Use a frontier reasoning model when:
Section titled “Use a frontier reasoning model when:”- the task spans many files
- the change has architectural consequences
- you need the model to recover from failures and keep a plan straight
Use a fast model when:
Section titled “Use a fast model when:”- you are iterating quickly
- the task is local and well-scoped
- autocomplete quality matters more than deep planning
Use a local model when:
Section titled “Use a local model when:”- your data cannot leave your environment
- you need predictable operational boundaries
- you can accept some capability tradeoffs for control
Keep One Eye on Access Models
Section titled “Keep One Eye on Access Models”Two tools can expose the same model through very different operating constraints.
- direct provider access
- aggregator access such as OpenRouter
- cloud-platform access such as Bedrock, Vertex, or Azure OpenAI
- local model serving via Ollama, LM Studio, or vLLM
This matters because provider choice changes retention policy, logging surface, and enterprise deployment options.
Where to Check Live Data
Section titled “Where to Check Live Data”- Artificial Analysis
- SWE-bench
- Aider Leaderboards
- official provider docs for current access details
Bottom Line
Section titled “Bottom Line”Do not choose a model from a stale table. Choose it from the workflow you need to support, then verify the live benchmark and access details before committing.