Why This Matters

Models get deprecated (Sept 2025 changelog). Hardcoding model names breaks systems when deprecations hit. Designing for churn is essential for production reliability.

Concept Explained

Deprecation-resistant architecture: abstract model choice behind an interface. Use model-agnostic prompts. Maintain a fallback matrix (if X deprecated, use Y). Plan migrations proactively.

How It Works

Don't hardcode model names in code. Use a config: model = env or config file. Build prompts that work across models. Maintain a fallback matrix. Track changelog for deprecations.

Step-by-Step Tutorial

1. Abstract model choice

2. Use model-agnostic prompts

3. Build fallback matrix

4. Track changelog

5. Plan migrations

Real-World Example

A team hardcoded 'claude-opus' in their CI bot. Sept 2025: claude-opus deprecated. CI broke. Emergency fix: update config to 'claude-sonnet'. Lesson: never hardcode model names.

Example Prompts / Commands / Code

# config/copilot.json
{
  "default_model": "claude",
  "fallback_model": "gpt-5",
  "task_overrides": {
    "refactor": "claude",
    "tests": "gpt-5",
    "quick_question": "gpt-5-mini"
  }
}

# Code uses config, not hardcoded names:
model = config['task_overrides'].get(task_type, config['default_model'])
copilot.run(prompt, model=model)

Deprecated model     Replacement       Test status
------------------------------------------------
claude-opus          claude-sonnet     Tested, OK
gpt-4-turbo          gpt-5              Tested, better
gemini-1.5-pro      gemini-2.5-pro     Testing

Common Mistakes

• Hardcoding model names — breaks on deprecation.
• No fallback matrix — scramble on deprecation.
• Not tracking changelog — surprised by deprecations.
• Model-specific prompts — don't port between models.

Best Practices

• Never hardcode model names.
• Use config-driven model selection.
• Build model-agnostic prompts.
• Maintain a fallback matrix.
• Track changelog monthly for deprecations.

Troubleshooting

Practical Exercise

Professional Challenge

Frequently Asked Questions

Further Reading

• Model deprecation changelog

Official References

• Changelog

Related lessons: IN-27, PR-04