Two Big Changes at Once
Anthropic has made two simultaneous announcements that affect anyone building on Claude: the deprecation of two models AND a significant change to the billing structure. In this video I break down both, and tell you exactly what to do today if you are affected.
Change 1 — Model Deprecations
Anthropic has deprecated two Claude models that were previously available via API. Applications calling these deprecated model IDs will either fail or be automatically routed to a successor model — depending on the endpoint. You need to know which you are on.
What to do:
- Audit every place in your codebase where a Claude model ID is hardcoded
- Check Anthropic's model deprecation documentation for exact sunset dates and successor models
- Update model IDs to current Claude versions (Claude 3.5 Sonnet or Claude 3 Haiku depending on your use case)
- Test thoroughly after updating — successor models may behave differently on edge cases in your prompts
Change 2 — Billing Split
Anthropic has changed how input and output tokens are priced. Previously, the pricing was relatively symmetric. The new structure applies a different multiplier to output tokens — meaning applications that generate long responses are now more expensive, while applications that primarily read/classify short inputs are cheaper.
Who is most affected:
- Applications generating long-form content, reports, or code — costs increase
- Classification, extraction, and routing applications — costs decrease
- RAG applications with short answers — likely cheaper overall
- Summarisation with long outputs — more expensive
Immediate Actions
- Today — run a cost projection using your actual token usage split (input vs output) against the new pricing
- This week — update any hardcoded model IDs and deploy to non-production for testing
- This month — review system prompts and response length instructions. If you are on a high-output pattern, explicit instructions to be concise can meaningfully reduce cost
- Ongoing — implement model and cost abstraction so future pricing changes require a config update, not a code change
The Pattern to Build Against
This is the third pricing or model change from a major AI provider in 90 days. The message is clear: AI model economics are not stable yet. Build your applications with a configuration-driven model selection layer. The cost of building that abstraction once is far less than the scramble of emergency migrations every time a provider changes their pricing or deprecates a model.
