Claude Opus 4.7: tokenization matters in a migration

One model, a different text split

In this retrospective April review, Claude Opus 4.7, announced on April 16, 2026, introduces a different tokenizer: the same text can produce more tokens. A tokenizer is the mechanism that divides the text processed by a model. Anthropic announcement.

Compare a complete task

For a coding agent, comparing only the advertised token price leaves part of the question unanswered. A useful protocol repeats the same requests, repository and tools, then observes total cost and verified results. An accepted fix, a partial response and lengthy reasoning without a solution should remain distinct outcomes in that assessment.

Review existing instructions

Migration is also an opportunity to examine accumulated instructions: contradictions, repeated requests and unclear stopping criteria make comparisons harder. A small selection of representative tasks, with execution records kept, is a practical starting point. The vendor’s published evaluations can guide exploration; they do not replace measurement on the work actually assigned to the agent.