Skip to content
Trazum

2.7.3

Released 2026-11-05

Fixed#

  • OpenAI's usage report left every Flex and Fast row out of the bill. from-openai, bill and the web Bill tab refused any row at a service tier but default, because until 2.7.2 there was no rate to price it at. There is now: the catalogue carries Flex, Fast and Ultrafast tables for 25 models. A row at one of those tiers, on a model whose row has that tier's table, is carried with service_tier and priced from the table exactly as a logged response naming the same tier is. A row whose model has no table for its tier (GPT-4o has no Flex), or a tier value Trazum does not know, is still left out and named, because pricing it at the standard rate would put a Flex bill at twice its cost. from-openai now says how many rows were priced at which tier.
  • Anthropic's usage report read fast-mode rows at the standard rate. The report says which speed ran when it is grouped by speed, and from-anthropic ignored the field, so a row of Opus 5 fast mode ($10/$50) was billed at $5/$25. A fast row is now written with usage.speed: "fast" and priced from the model's own fast rate, as a logged response carrying the field is; one on a model with no fast rate is left out and named. A report not grouped by speed that holds a model with a fast mode now says so, and the curl in docs/commands.md asks for the grouping.
  • US-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, on top of fast mode, caching and the long band, and every response and every Claude Code transcript line records the geography that ran in usage.inference_geo. Trazum dropped the field everywhere. The catalogue now carries the published multiplier on the 13 rows it applies to, and a call whose log says "us" is priced at it: through bill and every converter, the gateway (from the request's own inference_geo before the call and the response's after it), and from-anthropic, whose curl now asks for group_by[]=inference_geo and which says when a report never named a geography. "global", "not_available" and a log without the field keep the standard price. A test fails a Claude 4.6 or later row that arrives without the multiplier, and the weekly price check reads the page's sentence and fails if the multiple or the generation it starts at changes. A --pricing overlay accepts inferenceGeoMultipliers too, validated like every other rate, so a catalogue row copied into one is still valid.
  • trazum connect summed every tier into one row and billed it as standard. It asked each usage report for group_by[]=model alone, so Anthropic batch and fast mode, US-only inference, and OpenAI batch, Flex and Fast all arrived folded into one row per model and day, at the standard rate. The Anthropic pull now also groups by service tier, speed and inference geography, and the OpenAI pull by batch and service tier. A fast, US-only, Flex, Fast or Ultrafast row is priced from its own published rate, as bill prices the same calls from a log; a row no catalogue rate is the rate for is left out and named with its model, tier and token count. A store kept with connect --store converges across the change: the newest pull of a window and model replaces every older record of it, whatever its grouping, so a day stored before and re-pulled after is split, not doubled.

Release mechanics#

  • 2.7.2 published all four packages and then had no tag and no release page. GitHub refuses the workflow's token any ref whose .github/workflows/ files differ from main's, because GITHUB_TOKEN can never hold the workflows permission. A pull request that changed a comment in ci.yml merged while the Release run was publishing, so the tag on the release commit was refused; through the releases API that read as a bare HTTP 403. The step now pushes the tag over git, where the refusal names the file, and on refusal prints the one command that recovers: push the tag by hand, and the tag-triggered run creates the release without uploading anything. docs/releasing.md adds the rule that prevents it (don't merge a workflow change during a Release run), and backfill-releases gains version, commit and latest inputs for a missing release whose commit still matches main's workflows. A first version of this fix blamed the timing alone and was proven wrong by its own backfill run, which GitHub refused for the same reason; v2.7.2's tag needs a person. The MCP registry listing was skipped with it and stays at 2.7.1 until the next release.

Guarded#

  • The weekly price check now reads each provider's lifecycle too. Anthropic's deprecations page has one table of every model's state, and each OpenAI model page opens with a Status: line once the model is deprecated or shut down, and gives its context window. A model the provider has deprecated, retired or shut down that the catalogue still recommends is now a finding, and so is a context window that differs from the page; both fail the scheduled CI run. The catalogue had recommended GPT-5, mini and nano for weeks after their pages said "Status: Deprecated", and nothing compared the two. Planted on saved copies of the pages, a deprecation on each side and a changed window were each reported.