Codex logs cumulative token counts, which is what makes per-turn tier pricing possible at all. agent-cost takes the increase since the previous count and prices that increase using the settings recorded for that turn [7]. A session that opens on Standard, runs one turn on Fast, then returns to Standard therefore gets priced in pieces. Applying the session's final setting to the whole session would have charged the Fast turn at Standard rates [8].
One interval survives that rule. When the setting changes while a turn is still running, the reader cannot safely assign those tokens to the old or the new setting, so they stay unpriced [9].
The spread is what makes a default expensive. On the fixture, six million tokens estimate at $122 Standard and $305 Fast [4][5], a ratio of 2.5 and a gap of $183 [1][2]. Per million tokens that is $20.33 against $50.83 [3]. A reader that quietly assigned unknown usage to Standard would understate a Fast session by 60 percent of its own estimate [4], and the author's stated reason for refusing to do it is that the number would come from settings the reader could not establish [21].
For $122 to describe anyone's spend, several things have to hold. The mix has to be the fixture's: two million uncached input, two million cached input, two million output tokens [3]. The model ID has to be exactly gpt-6-astra, and the usage has to fall at or after September 5, 2026, 17:18:23 UTC, or it stays unpriced even when the mode is identifiable [10]. A catalog boundary specified to the second is more rigour than most price tables bother with. And the bundled catalog has to match what is actually billed, which the write-up does not claim: the table is a test fixture, not a list of current official prices [2].
Instead of a bare amount, the output carries pricing_status and a separate unpriced_tokens count [11]. The unknown row consequently keeps all six million tokens visible as unpriced [12], and a zero there means none of the usage could be priced, not that it was free [6]. Dropping the row would have hidden that the usage happened; keeping only the zero would have made it look free [22].
Two caveats in the notes deserve the same weight as the fix. The `priced` state means the usage could be calculated against the selected catalog, not that a charge was confirmed [13]. And the whole determination rests on request settings recorded in the log, so it does not establish the processing mode the server used or what was billed [14]. The field names came from reading the public source for Codex 0.153.4 [15]; the reader checks that a service_tier setting belongs to the same session and matches the target turn and model before using it, with `default` mapping to Standard and `priority` to Fast [16].
The Fast-priced-as-Standard bug never reached users. Version 0.1.0 shipped without Astra rates at all, so it left Astra usage unpriced, and the mispricing was caught while Astra support was being built [18]. Scope stays narrow by the author's own statement: missing history, concurrent updates, and alternative storage formats are not covered [19].
