Skip to content
Trazum

2.7.1

Released 2026-11-05

Added#

  • Sixteen earlier OpenAI models are priced, each from its own page: GPT-5.5, GPT-5.5 pro, GPT-5.4, GPT-5.4 pro, mini and nano, GPT-5.2 and GPT-5.2 pro, GPT-5.1, GPT-5 pro, GPT-4.1 and mini, GPT-4o and mini, o3 and o3-pro, with their dated snapshots. Every call to one of them read as an unpriced model. The ones OpenAI marks deprecated are priced for the logs that hold them and never recommended. GPT-5.5 and 5.4 (and their pro models) bill a prompt over 272K at 2x input and 1.5x output, per request. gpt-4o-2024-05-13 is shut down and was priced differently, so it is not folded into GPT-4o.
  • GPT-6.1 Terra: $1/$5, cache read 5% of input, the same 272K band.
  • Kimi K3.1 Pro ($2/$10, cache read $0.20, writes $2 and $4) and its high-speed variant ($4/$20). Kimi K3 gains its batch price (60%), which the batch page now lists. The K2 series is closed to accounts registered from 2026-11-01 and retires by the end of December, so it is priced and no longer recommended. Kimi's Markdown pages now ship their price tables empty; the figures were read from the HTML page's compiled tables.
  • The November price review. Anthropic, OpenAI, DeepSeek, Moonshot, xAI and Mistral were re-read from their own pages on 2026-11-05; every price already in the catalogue matched, and each provider's review date moves. Google could not be re-read: its Gemini API docs, pricing included, now live on aistudio.google.com, which this environment does not reach. Its date stays at 2026-10-10 rather than move without a read, so the CLI will warn about Google prices from 2026-11-25 until someone reads them.

Fixed#

  • OpenAI's 12-hour cache writes were priced as 30-minute ones: half their cost. GPT-5.6 and later now sell a 12-hour cache entry in the Responses API at 2.5x input, against 1.25x for the default 30 minutes, and report the two apart in input_tokens_details.cache_write_tokens_details (ttl_30m_tokens, ttl_12h_tokens). The split is read, each lifetime is priced at its own rate, and the TTL fit names the 12-hour entry as the one to buy when turns are further apart than 30 minutes. An unsplit write on an OpenAI model is the 30-minute kind, because OpenAI returns the split whenever a 12-hour write is possible; Kimi uses the same field without that rule and is still said to be assumed.

  • The gateway billed OpenAI cache writes as plain input. On GPT-5.6 and later a write costs 1.25x input, and prompt_tokens includes it; the gateway left it there at 1x. Writes are now taken out of the prompt and priced as writes, in whole and streamed responses alike.

  • Anthropic fast mode was priced at the standard rate: half its cost. Fast mode (speed: "fast") bills Claude 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 that ran in usage.speed. Trazum dropped the field, so a fast call read from a Claude Code transcript, a logged API response or the gateway was billed at the standard rate. The field is now kept end to end and priced by a new speed tier condition, with the cache prices on top of the fast rate as Anthropic's page says. A log that does not record a speed is priced as standard and is not marked uncertain: fast mode is an opt-in every response names. The gateway also prices a request that asks for fast mode at the fast rate before sending it, so a budget judges the real cost. Read from platform.claude.com's pricing and fast-mode pages on 2026-11-05.

  • Claude Sonnet 4.5 was unpriced. It is deprecated and still served until 2026-11-30, so it is in most logs of the past year, and every call to it read as an unpriced model. It is priced now from its own model page ($3/$15, 200K context), never recommended. The retired Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are on the same pricing page and are not priced yet: a row also needs a context window, from pages not read today.

  • A usage file past about 512 MB no longer takes the whole bill down. JavaScript cannot hold a longer string, so reading such a file whole failed with the runtime's "Invalid string length", naming no file, and nothing else was billed either. A Claude Code session that read large files reaches that size: its transcript holds every tool result. Files over 256 MiB are now read line by line with their message text set aside (slimUsageLine in @trazum/core), which every converter ignores anyway, so the bill is identical: the tests force every file down this path and compare the receipts byte for byte. Gzipped logs that inflate past the line take the same path. A file still too large after that is refused by name, with what to do. The web app's Bill tab had the same ceiling in File.text() and now reads such a file the same way, streamed in the browser; a file it still cannot hold is named under the bill rather than failing the tab. Measured on this machine: a 617 MB session billed in 6 s at under 1 GB of memory; 321 MB across 400 sessions, 40,000 calls, in under 4 s. See docs/usage-logs.md.