Repository navigation
v2.7.3
2.7.3 — Every tier the provider names, billed at its own rate
Adoption, read 2026-11-05: downloads of @trazum/cli in the last 30 days: unavailable (fetch failed), 1 GitHub stars, MCP registry latest 2.7.1.
A patch of corrections to what a bill charges, all of the same kind: the
provider said which tier, speed or geography a call ran at, and Trazum read it
as standard. If any of your Claude workspaces pins inference to the US, if
you use fast mode, or if you bill OpenAI's or Anthropic's usage report or
trazum connect, run it again.
npx -y @trazum/cli@2.7.3 billUS-only inference was priced a tenth under its cost. Anthropic bills
inference_geo: "us" at 1.1x on every token category of Claude 4.6 and later
(input, output, cache writes and reads), on top of fast mode and Haiku 5.5's
long band. Every response and every Claude Code transcript line names the
geography that ran in usage.inference_geo, and Trazum dropped it. The
catalogue now carries the published multiplier on the 13 rows it applies to,
and bill, every converter, the gateway and from-anthropic price a US-only
call at it. A global call, or a log without the field, is unchanged.
OpenAI's usage report left every Flex and Fast row out of the bill.
from-openai, bill and the web Bill tab refused a row at any tier but
default. A row at Flex, Fast or Ultrafast is now priced from that tier's own
table, as a logged response naming it is; a tier with no table on that model
(GPT-4o has no Flex) is still left out and named.
Anthropic's usage report read fast-mode rows at the standard rate. A row
of Opus 5 fast mode ($10/$50) was billed at $5/$25. from-anthropic now
prices it from the model's fast rate, and says when a report was never asked
its speed or geography.
trazum connect billed every tier as standard. It asked each usage report
for the model alone, so batch, fast mode, US-only inference, Flex and Fast all
arrived as one row per model and day. It now asks for the groupings that change
a price and prices each row by them; a row no catalogue rate covers is left out
and named with its token count. A store kept with connect --store converges
across the change: a day stored before and re-pulled after is split, not
doubled.
The weekly price check reads the data residency sentence too. If the
multiple or the Claude generation it starts at moves on Anthropic's pricing
page, the scheduled run fails, and a test fails any Claude 4.6 or later model
added without the multiplier.
One thing still waits on a person. 2.7.2 is on npm, but its tag and
GitHub release page are not: GitHub refused the workflow's token the tag
because a pull request changing .github/workflows/ merged while that run was
publishing. docs/releasing.md names the rule and the one command that
recovers it.
No breaking change. All packages, the plugin and MCP discovery advance
together to 2.7.3.