Repository navigation
v2.7.1
2.7.1 — Three bills that were too low, and a file too big to read
Adoption, read 2026-11-05: downloads of @trazum/cli in the last 30 days: unavailable (fetch failed), GitHub stars: unavailable (HTTP 403), MCP registry latest 2.7.0.
A patch of corrections. Three of them were in the flattering direction: the
bill came out lower than what the provider charged. If you bill Claude Opus
in fast mode, or OpenAI with 12-hour caching, run it again.
npx -y @trazum/cli@2.7.1 billAnthropic fast mode was billed at half its cost. Fast mode bills Opus 5.5
at $8/$40 and Opus 5 and Opus 4.8 at $10/$50, twice their standard rates, and
every response records the speed in usage.speed. Trazum dropped the field.
Now Claude Code transcripts, logged API responses and the gateway all keep it
and price by it. A log with no speed is priced as standard, because fast mode
is an opt-in that every response names.
OpenAI's 12-hour cache writes were billed at half their cost. GPT-5.6 and
later now sell a 12-hour cache entry at 2.5x input, beside the 30-minute one
at 1.25x, and report them apart. Both are read and priced at their own rate.
The gateway, separately, had been billing OpenAI cache writes as plain input;
it no longer does.
A session past about 512 MB took the whole bill down. JavaScript cannot
hold a longer string, and a Claude Code session that read large files grows
past it. Files over 256 MiB are now read line by line with their message text
set aside, which nothing is billed by. The tests prove the bill is the same,
byte for byte. The web app's Bill tab does the same in the browser. Measured:
a 617 MB session in 6 seconds.
Twenty-one models that read as unpriced now have prices, each from its
provider's own page:
- Claude Sonnet 4.5 (deprecated, served until 2026-11-30);
- GPT-6.1 Terra, and sixteen earlier OpenAI models still in logs: GPT-5.5,
5.4, 5.2, 5.1 and their pro, mini and nano siblings, GPT-5 pro, GPT-4.1,
GPT-4o and o3; - Kimi K3.1 Pro and its high-speed variant.
Kimi K3 gains its batch price. Deprecated models, and the Kimi K2 series that
retires in December, are priced and never recommended.
The November price review. Anthropic, OpenAI, DeepSeek, Moonshot, xAI and
Mistral were re-read on 2026-11-05 and every existing price matched. Google
could not be read: its docs moved to a domain this build environment cannot
reach. Its prices stay as read on 2026-10-10, and the CLI will say they are
stale from 2026-11-25 until someone reads them.
No breaking change. All packages, the plugin and MCP discovery advance
together to 2.7.1.