Author here. I built ccusage_go (https://github.com/SDpower/ccusage_go) to parse Claude Code's local JSONL session transcripts after noticing instruction-following degrade in long sessions.
Data showed 97.7% cache overhead vs 2.3% actual compute. Dug into why — caching is structurally optimal for text-only, which explains both the product roadmap and the 2026 hiring push for vision/audio engineers doing work competitors finished in 2023.
Related writeups with the raw data:
https://blog.sd.idv.tw/en/posts/2026-03-25_claude-code-cache... https://blog.sd.idv.tw/en/posts/2026-03-25_claude-code-10-ti...