Cheap Model for Testing / Expensive Model for Production

Core Rule

When writing, editing, or reviewing code that calls an LLM or any pay-per-usage API (OpenAI, Anthropic, Google AI, Cohere, Mistral, Replicate, AWS Bedrock, Azure OpenAI, etc.), always default to the cheapest available model for development and testing purposes.

What To Do

During development and testing

• •Always choose the cheapest model available for the provider being used.
• •Common cheap model choices (use the latest available version):
  • •OpenAI: gpt-4o-mini (or gpt-3.5-turbo if mini is unavailable)
  • •Anthropic: claude-haiku-4-5-20251001 (or the latest Haiku variant)
  • •Google AI / Vertex: gemini-2.0-flash (or the latest Flash variant)
  • •Mistral: mistral-small-latest
  • •Cohere: command-r (not command-r-plus)
  • •AWS Bedrock / Azure OpenAI: whichever is the cheapest equivalent of the above
• •If the user has already specified a model and it is NOT the cheapest, proactively suggest switching to the cheapest one for testing and explain why (cost savings during development).
• •If a config file, environment variable, or constant defines the model, set it to the cheap option and leave a code comment like:
  code

  # TODO: Switch to production model before deploying (e.g., claude-sonnet-4-5-20250514)
  

When deploying or finalizing for production

• •Before changing any model to a more expensive/capable one, always ask the user:

  "This code is using [cheap model] for testing. Are you ready to switch to a production-grade model for deployment? If so, which model would you prefer?"

• •Never silently upgrade to an expensive model. Always get explicit confirmation.
• •Suggest sensible production model options for the provider in use (e.g., claude-sonnet-4-5-20250514, gpt-4o, gemini-2.0-pro).

How To Detect Deployment Context

Consider the task to be "deployment" or "production-ready" when the user says things like:

• •"deploy", "ship it", "push to production", "release", "go live"
• •"finalize", "production-ready", "ready for launch"
• •"switch to the real model", "use the good model now"
• •Creating a PR described as production/release-ready

In any of these cases, stop and ask before changing the model.

Summary