ClaudeBar v1.3.2

Both cost columns were wrong. Claude overstated by roughly 7x, Codex understated by roughly 3x.

Claude

• One response was billed once per content block. Claude Code writes a line per block and repeats the response's usage on each, so a four-block answer counted four times. Now deduplicated per API call, keeping the final line - the mid-stream ones carry a placeholder token count.
• Advisor turns were partly invisible. A second model runs inside the response and its tokens appear only in usage.iterations; those are now read and billed at the model each names.
• The pricing table was two generations stale, charging Opus at $15/$75 per Mtok when it is $5/$25. It now carries the current models, prefers the longest matching prefix, and falls back within the model family.

Codex

• Resumed sessions lost everything before the resume. codex resume restarts the cumulative counter at zero in the same file; earlier epochs are now banked, with a guard so a resume that re-derives its total slightly downwards is not mistaken for a restart.
• Sessions over 100 MB were skipped entirely. Those files are large because of base64 payloads, not line count, and are now included.
• Cached and reasoning tokens were billed twice. Codex reports them inside input_tokens and output_tokens; each is now charged once at its own rate.

Known limitation: a session is bucketed to the day it started, so the daily figure can drift across midnight. Monthly totals are unaffected.

The Windows installer below is signed and timestamped.