A patch with no new command. The one line works on a first run:
npx @trazum/cli bill with no path reads Claude Code's own folder, labels each
session by project, and prints a statement on a terminal. Every call is billed
at its own moment: a report used to price every call at the date it ran, so a
clock-priced log read at peak was all peak and a promotion vanished once it
ended; and a row can now carry the prices it had before a published change,
which DeepSeek's 2026-09-10 cut needed. Google, DeepSeek, Mistral and OpenAI
were re-read: DeepSeek had cut V4 Flash's price, the cache figures were wrong
on Google, Mistral and DeepSeek, and the ids OpenAI's own usage records carry
now price. The web app opens
on the bill, and the npm listing names Claude Code.
Changed#
OpenAI, re-read on 2026-10-11: the ids OpenAI's usage records carry now price. GPT-5, mini and nano matched the pricing page to the cent. Their dated snapshots (
gpt-5-2025-08-07and the mini and nano ones), which is what an OpenAI usage export names, were unpriced gaps; they are declared as aliases now, read off the deprecations page that retires them on 2026-12-11. The section comment said a cached read costs "about half the input price"; it costs a tenth, as the rows already said. The 1,024 cache minimum is what OpenAI publishes for GPT-5.6 and later, and its guide says the floor "varies by request settings" for these older models, which the notes now say. The current generation is not added: gpt-6-astra, gpt-6.1-sol, gpt-6-sol, gpt-6-luna and gpt-5.6 sol, terra and luna all double their rates for prompts over 272K tokens, a per-request condition no log Trazum prices passes yet.Mistral caches, and this table said it did not. Re-read on 2026-10-11 from mistral.ai/cloud/inference and the Large 3, Small 4 and Medium 3.5 model cards, which carry their prices in the page markup and bind the
-latestids. Every price matched. The inference page offers "Cached input tokens -90% on input token" and every card marks its input price as carrying that discount, so a Mistral log's cache reads are now priced at 10% of input rather than at the full rate, and the notes no longer say reordering saves nothing. How caching starts and its minimum are on docs.mistral.ai, which is not reachable from here, so the three rows carrycaching: 'unknown'and no minimum, and Trazum gives no caching advice for them rather than inventing either. The MCP model list says "caches, minimum unpublished" for them instead of "no caching". Mistral Large 4 ($1.36/$4.18) is on the page and not yet added: its card states no id.DeepSeek, re-read on 2026-10-10: V4 Flash was retired and its price cut, so one id billed two ways in one month. DeepSeek released V4.1 Flash on 2026-09-10 and announced "New pricing takes effect at 04:00 UTC on Sept 10, 2026": $0.15/$0.60 off-peak and double at peak, a cache hit at 2% of input. The old
deepseek-v4-flashid is still accepted and now bills at that price. The catalogue had one price for all time, $0.22 and $0.66, so every call since the cut was priced about 47% too high on input. A row can now carry its earlier prices (history, each period with the instant it ended, its own tiers and its own cache ratio), anddeepseek-v4-flashcarries the old price until 04:00 UTC on 2026-09-10 and the new one after it, to the millisecond.deepseek-flashis added at the new price. V4 Pro matched the page unchanged. The flat price before clock pricing began (16:00 UTC on 2026-08-16) is on no page this table can read, so calls before then are priced at the first period and the row says so.The Google catalogue was re-read on 2026-10-10, and three things about its caching were wrong. Every price matched. What did not: implicit caching is on by default for every Gemini 2.5 and newer model, so caching is
automaticacross the board rather than a prefix to mark; the minimum is 4,096 tokens on 3.6 Flash (not 1,024) and 2,048 on 2.5 Flash (not 1,024); and the 2.5 models read the cache at 10% of input, not the 25% this table carried, so a 2.5 log's cache reads were priced at two and a half times what the page charges today. Gemini 3.8 Flash is added (Google's newest Flash, the same $0.75/$3.75 introductory rate as 3.6 through 2026-12-31). Gemini 3.1 Pro Preview stays out, and 2.5 Pro's prompts over 200k stay at the lower rate with a note saying so: both are priced by prompt length per request, which no log Trazum prices carries yet. The README's cache rows say the new ratios and minimums.The npm listing of
@trazum/clisays what it is for. Its description was about a prompt bill and its keywords wereprompt, llm, tokens, cost, cli, ci: nothing a person searching npm for what Claude Code cost them would type. It now names Claude Code, the sources it prices and the PDF, and its keywords includeclaude-code,anthropic,openai,usage,billingandfinops.trazum billmeets a first run where it is. Three places the README's one line stopped short, each measured on a real~/.claude/projects:- No path was a refusal. It now reads Claude Code's own folder,
$CLAUDE_CONFIG_DIR/projects,~/.claude/projectsand~/.config/claude/projects, every one that exists, naming each; with none there it names every place it looked. - Sessions had no project.
from-claude-codehas labelled a directory by its project folders since 1.77, andbillread the same transcripts with no label, so the bill could not say which project spent the money. Same default now, with--no-label-from-projectto decline and--labelto set one. - A terminal got the JSON receipt. On a TTY
billnow prints a statement: the total, a table by project, a table by model, and what is not in the total. A pipe still gets the receipt, so scripts see no change;--receiptand--tableforce either. A folder of hundreds of transcripts is one line per shape on stderr rather than one per file.
- No path was a refusal. It now reads Claude Code's own folder,
The web app opens on the bill. Since 2.4.0 the README, the landing page and the CLI all open on one thing,
trazum bill, and the web app opened on the prompt optimiser: a visitor sent by the line they had just read landed on a different product. The bill is now the default panel, the Measure group sits first on the rail, a bare/drops?tab=billfrom the address, and the guided tour starts on the bill. Every panel is where it was, one click away, and?tab=optimisestill opens the optimiser. The deep-link guard that pinned the old default now pins this one, as the product decision it says it is.The README's first section shows the PDF line and
npx skills add.--pdf-out bill.pdfsits under the one-line start, and an agent that reads Agent Skills from a repository is told it can install the skill withnpx skills add Davmunrey/Trazum, which lists it today.
Fixed#
A bill priced every call at the date the report ran, not the date of the call. A DeepSeek log read during peak hours came out entirely at peak, and the same log read on a Saturday entirely off-peak; a call made inside a promotion lost it once the promotion had ended.
profileUsagepassed one date to every record. Each call is now priced at its own timestamp, and the report's date applies only to a record with no clock and to the forward-looking findings. Tests read one peak and one off-peak call at a peak hour and on a weekend and get $1.98 both times; before this it was $2.64 and $1.32.The packaged Action pin in
README.mdanddocs/running.mdadvanced to 2.4.2's release commit, the onev2.4.2points at, in the pull request after the release asdocs/releasing.mdsays it must.The guided tour's welcome said nothing pasted is uploaded. The bill is read in the page; Optimise, Write and Compare post the prompt to the server. The sentence now says which is which, in English and Spanish, and
copy-claims.test.mjsholds it.