Claude Code Opus 4.7 Token Surge

In Claude Code, Opus 4.7 processes independent Read operations one file per request instead of batching them together, causing a surge in token usage.

When documenting a small Express/SQLite API repository (12 files, about 500 LOC), saved Claude Code JSONL transcripts showed that 4.6 batched things into a handful of requests, while 4.7 repeatedly performed single-file reads only.

The key observations are as follows.

• Opus 4.6: 3-6 requests, centered on batch processing
• Opus 4.7: 16-20 requests, mostly 1 Read per request
• Cache read tokens increased from 50,566 / 80,111 to 432,557 / 464,258
• Average cache read tokens per request also rose from 16.9k / 13.4k to 27.0k / 23.2k
• Total tokens also expanded from roughly 73k / 106k levels to 454k / 528k levels

The author explained that this difference is not an actual context explosion, but the result of repeatedly reusing the same large Claude Code system context across multiple model requests. In other words, this reads less like a performance degradation and more like a possible regression that directly impacts token cost and usage limits.