Three AI API changes that quietly break existing code

A DEV Community post details three API changes that break shipped code without loud errors: Claude Haiku 5.5 replacing 4.5 with a new tokenizer and no budget_tokens, stricter allowed_hosts in Claude Managed Agents, and GPT-6 splitting into three models with a 100x price spread.

• Haiku 5.5 replaced 4.5: budget_tokens is gone, replaced by adaptive thinking and output_config.effort
• Haiku 5.5's new tokenizer makes the same text about 30% more tokens
• In Claude Managed Agents, allowed_hosts now gates web_search and web_fetch, otherwise url_not_allowed
• GPT-6 spans Luna ($0.10/$0.50), Sol ($2/$10) and Astra ($10/$50) per million tokens