v2.5.0

2.5.0 — Each request at its own size

Adoption, read 2026-10-11: downloads of @trazum/cli in the last 30 days: unavailable (fetch failed), GitHub stars: unavailable (HTTP 403), MCP registry latest 2.4.3.

A minor, and no new command. Two published formats grew a field each: a usage
row may say how many requests it sums, and a receipt can name a range it could
not close.

Priced by the size of each request

npx @trazum/cli bill

Four providers now charge some models by prompt length, per request: Claude
Haiku 5.5 is dearer over 100,000 input tokens, Gemini 3.1 Pro Preview over
200,000, OpenAI's GPT-5.6 and later over 272,000, and every current Grok from
200,000 exactly. A request over the line pays the higher rate on all of it,
cache included. Until this release those models were left out rather than
priced at a figure up to five times wrong.

• A call in a log is billed in its own band: a Claude Code transcript, an
  OpenTelemetry span, a LiteLLM or Helicone row.
• A row that sums requests (a usage report's hourly bucket, an OpenRouter
  day) cannot say how large any one prompt was. It is billed at the dearer rate,
  and every surface prints what the same tokens cost at the cheaper one: "$X of
  this total, $Y at the cheaper rate", in bill, profile, the receipt (a
  tier-ceiling gap), connect, the exported documents, the MCP report and the
  web app. A row whose whole prompt total is under the line is decided.
• A usage row can say how many requests it sums, as requests. The
  converters for usage reports write it for you.

Current models priced

• OpenAI: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol,
  Terra and Luna. Cache writes cost 1.25x input and an entry lives at least 30
  minutes, so the cache TTL verdict now judges a pause against the model's own
  entry rather than Anthropic's five minutes.
• Anthropic: Claude Haiku 5.5, $0.10/$0.50 up to 100,000 tokens.
• Google: Gemini 3.1 Pro Preview, $2/$12 up to 200,000.
• Moonshot: Kimi K3 ($3/$15, writes billed by TTL like Anthropic), K2.7
  Code and its high-speed variant, K2.6. kimi-k2 was discontinued on
  2026-05-25: still priced, no longer recommended.
• xAI: Grok 4.7, 4.6, 4.5, 4.3, the 4.20 family and Build 0.1.

Fixed, and what it was

• A Responses API log billed every cached token at the full input rate.
  OpenAI's default API reports usage as input_tokens with the cached part in
  input_tokens_details, and only the Chat Completions spelling was read. A
  GPT-5 call reading 90% of its prompt from cache came out at $0.135 instead of
  $0.034, and the cache verdict said caching was never used.
• grok-4 after 2026-05-15 was billed at $3/$15. xAI retired it that day
  and has billed it at Grok 4.3's $1.25/$2.50 since. Both periods are priced
  now, so a log spanning the date is billed both ways.
• OpenAI's usage report billed cache writes as plain input, through
  from-openai and connect.
• The findings under a bill priced their calls at the report's date, not
  the call's; connect billed every bucket the same way; a budget priced every
  model with one set of cache ratios.
• A block of Spanish messages had no accents.