← All Insights

Claude 3.5 Haiku, 3.7 Sonnet, GPT-4o: The Deprecation Wave Is Here

model-deprecationmigrationanthropicopenai

Claude 3.5 Haiku, Claude 3.7 Sonnet, and GPT-4o are all entering deprecation within the same window. OpenAI retired GPT-4o from ChatGPT on February 13th. If you’re running production systems that reference specific model IDs, this is your migration notice.

Model deprecation is the maintenance cost nobody budgets for. You build a system, tune prompts for a specific model’s behavior, validate outputs against expected patterns—then the model gets retired and you start over.

This isn’t a one-time event. It’s the new normal. Model generations are getting shorter. The gap between “cutting edge” and “deprecated” is measured in months, not years. GPT-4o was the flagship model less than a year ago. Claude 3.5 Haiku was the cost-efficient workhorse. Both are now being phased out.

The obvious thing that breaks is hardcoded model IDs in API calls. That’s fixable in an afternoon.

The less obvious thing that breaks is prompt behavior. Every model has quirks—formatting tendencies, instruction-following patterns, reasoning approaches. Prompts tuned for Claude 3.5 Haiku won’t perform identically on Claude 4.5 Haiku. Outputs you validated against GPT-4o’s behavior will drift on whatever replaces it.

If you have evaluation suites, now is when they earn their keep. If you don’t, you’re deploying model upgrades blind.

  1. Audit your model references. Search your codebase for specific model IDs. Every hardcoded string is a future migration task.
  2. Abstract model selection. Use configuration or environment variables, not inline strings. Make model swaps a config change, not a code change.
  3. Build evaluation pipelines. You need automated ways to verify that a model swap doesn’t break your specific use cases. Manual testing doesn’t scale across multiple deprecation cycles.
  4. Budget for it. Model migration is recurring maintenance. If your project plan doesn’t include “model deprecation response” as a line item, add it now.

The models will keep getting better. They’ll also keep getting retired. Build your systems to survive both.