GitHub issue #46917 in anthropics/claude-code, opened 2026-04-12 by Adrian-Mteam, records a report that Claude Code versions 2.1.100 and 2.1.101 consume approximately 20,000 more cache_creation_input_tokens per request than v2.1.98 despite sending fewer bytes in the client request payload, characterizing the difference as server-side token inflation; subsequent thread discussion records user verification and discussion of workarounds.
Verbatim Excerpts from Cited Records
“Claude Code versions 2.1.100 and 2.1.101 consume ~20,000 more `cache_creation_input_tokens` per request than v2.1.98, despite sending **fewer bytes** in the request payload. The inflation is server-side and version-specific (likely User-Agent routing).”
·
anthropics/claude-code (issue 46917 opening body, Summary section)
·
2026-04-12T08:14:34Z
·GitHub Issue ↗
“**v2.1.100 sends 978 fewer bytes than v2.1.98 but is billed 20,196 MORE tokens.** This rules out any client-side payload difference — the inflation happens server-side after the request is received.”
·
anthropics/claude-code (issue 46917 opening body, Evidence section)
·
2026-04-12T08:14:34Z
·GitHub Issue ↗
“The 20K extra tokens are `cache_creation_input_tokens` — this means they enter the model's context window, not just the billing ledger.”
·
anthropics/claude-code (issue 46917 opening body, Quality concern section)
·
2026-04-12T08:14:34Z
·GitHub Issue ↗
Counter-record: No in-thread refutation of the reported measurements as of 2026-10-04: 42 comments corroborate token inflation observations and discuss client User-Agent overrides as workarounds. Thread discussion notes that while proxy metrics demonstrate cache_creation inflation across tested versions, users cannot independently inspect hidden server-injected prompts without official provider disclosure.
Falsification probe: If issue #46917 is deleted without archive, if the author edits the opening body so the quoted sentences no longer match verbatim, or if the author/maintainers post a definitive retraction showing the proxy measurements were flawed or fabricated, this claim drops to Withdrawn. Ongoing thread discussion regarding version-specific fixes or prompt modifications does not invalidate the recorded measurement report unless the opening record is retracted.
Audit verification: Verified 100% continuous verbatim substring against official GitHub REST API by
scripts/verify_claim_verbatim.py.