JSONL usage.input_tokens and usage.output_tokens are unreliable upstream — ccusage undercounts significantly

[Magnus-Gille]

While building a Claude Code energy monitor, I discovered that the JSONL session logs that ccusage reads contain placeholder/incomplete token values. This is an upstream Claude Code bug (not a ccusage bug), but it means ccusage's reported totals are significantly undercounted.

What's wrong upstream

Claude Code's JSONL logs write usage values from early streaming events and never update them to final values:

• input_tokens: ~75% of entries have a value of 0 or 1 (streaming placeholder, never finalized). Undercount: 100-174x.
• output_tokens: Excludes thinking tokens (which Anthropic classifies as output tokens). Undercount: 10-17x on Opus.
• cache_read_input_tokens and cache_creation_input_tokens: Accurate (~1x vs ground truth).

Evidence

Comparing JSONL sums (deduplicated by requestId) against Claude Code's statusbar cumulative totals for the same day:

Metric	JSONL (dedup)	Statusbar	Ratio
Input tokens	41,444	7,199,162	174x
Output tokens	183,829	3,208,365	17x
Cache read	104,353,324	114,798,863	1.1x
Cache creation	3,170,696	2,717,775	0.9x

Cache metrics matching at ~1x confirms both sources track the same API calls — the discrepancy is in the recorded values, not missing entries.

Impact on ccusage

Cost and token reports from ccusage will be significantly lower than actual usage. For a heavy Opus day, ccusage reported ~$50 where statusbar-derived estimates put the equivalent API cost at ~$446.

Upstream issues

• anthropics/claude-code#28197 — our report (input + output)
• anthropics/claude-code#25941 — output placeholder (closed)
• anthropics/claude-code#21971 — output streaming incomplete
• anthropics/claude-code#27361 — missing message_stop event

Suggestion

Until this is fixed upstream, it might be worth adding a note/warning to ccusage output that input and output token counts may be understated due to known JSONL data quality issues. Full investigation: FINDINGS.md.