This article breaks down three unadvertised breaking changes in Anthropic’s Opus 4.7 release—budget_tokens now errors, the tokenizer uses 35% more tokens, and thinking tokens are hidden but still billed. It also highlights a sharp drop in long-context performance and user reports of increased hallucinations, and recommends updating code to use “effort”: “xhigh” and reviewing the migration guide.

• Opus 4.7's tokenizer uses ~35% more tokens for the same text, inflating costs and context budgets with no performance gain
• `budget_tokens` now throws a 400 error with no warning—must switch to "effort" levels (low/medium/high/xhigh/max)
• Thinking tokens default to hidden ("omitted") but are still billed, so you pay for blank output blocks
• Long-context retrieval accuracy on MRCR v2 crashed from 78.3% (at 1M tokens in 4.6) to 32.2% in 4.7, alongside user reports of increased hallucinations