Skip to content
Trazum

2.7.2

Released 2026-11-05

Fixed#

  • OpenAI's Flex calls were billed at twice their cost, and Fast calls at half or less. OpenAI sells Flex at about half the standard price, Fast (Priority until 2026-07-30) at up to twice and Ultrafast at six times, each from its own table on the pricing page, read 2026-11-05. The response names the tier that ran in its top-level service_tier, and Trazum ignored it, so every such call was priced at the standard rate. Now a logged Chat Completions or Responses object and the gateway, whole or streamed, keep the field, and the call is priced from that tier's table, long-context band included, for the twenty-five OpenAI models the tables list. "priority" is read as Fast, as OpenAI's guide says GPT-5.6 and earlier return it, and "default", which a Fast request downgraded under load also returns, is the standard rate. The gateway judges a budget by the tier the request asks for. If you use Flex or Fast and bill logged responses, run the bill again. The rates are written out per model because they are no single multiple: GPT-4o's Fast is 1.7x, GPT-5.2's 2x. A tier a model has no table for (GPT-4o has no Flex) prices at the standard rate, and OpenAI's usage report still leaves non-default tiers out and names them.

  • GPT-5, GPT-5 mini and GPT-5 nano were still recommended, and are deprecated. Each model page, read 2026-11-05, opens with "Status: Deprecated", and the catalogue still offered all three as cheaper alternatives, nano as the cheapest OpenAI model there is. They are priced for the logs that hold them and no longer recommended.

Added#

  • Four retired Claude models are priced: Claude Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5, from Anthropic's pricing, deprecations, context-windows and prompt-caching pages, read 2026-11-05. They are retired on the Claude API and still served on Google Cloud and Bedrock, and every Claude Code log from before their retirement carries them. Each of those calls read as an unpriced model, so a bill spanning them was low, and an Opus 4.1 session by a lot: it bills at $15/$75, three times Opus 4.5. Opus 4 bills the same, Sonnet 4 at $3/$15 and Haiku 3.5 at $0.80/$4, with the standard cache and batch multipliers, a 200k context window and the cache minimums the caching page gives. They are priced for the logs and never recommended. Claude Sonnet 3.7 and Haiku 3 are retired and no longer on the pricing page, so they stay unpriced rather than carry a figure no page publishes.

  • GPT-5.6 Cyber is priced: $12.50/$75, cached input $1.25, the 30-minute write $15.625 and the 12-hour one $31.25, a 400,000-token window with at most 272,000 input tokens, and no batch, from the pricing page and its own model page, read 2026-11-05. It needs Daybreak programme approval, so it is never recommended. The price-drift check below found it on its first run.

  • Codex sessions on the Codex models are billed. Codex CLI names gpt-5.3-codex, and before it gpt-5.2-codex, gpt-5.1-codex and its Max and Mini variants, gpt-5-codex and codex-mini-latest; none was priced, so every such call in a Codex bill read as unpriced. Each is now priced from its own model page, read 2026-11-05: $1.75/$14 for 5.3 and 5.2, $1.25/$10 for 5.1, 5.1 Max and 5, $0.25/$2 for 5.1 Mini, $1.50/$6 for codex-mini-latest (whose cache read is 0.25x, not 0.1x), no batch for any. chat-latest ($5/$30) is priced too. gpt-5.3-codex is deprecated and the rest are shut down, so all are priced for the logs and never recommended; chat-latest is not recommended because it moves with ChatGPT. gpt-5-search-api is on the pricing page with no model page, so it stays unpriced, recorded in the drift check with that reason.

  • gpt-5.6 is priced as GPT-5.6 Sol. Its model page says the alias routes there, and a log that names the alias read as unpriced.

Honest gaps, stated#

  • GPT-5.6 Sol's $4/$20 is promotional until 2026-11-30, its page says, as "a 20% reduction in input pricing and a 33% reduction in output pricing". The price after it is not published as a figure, and one derived from two rounded percentages would be a price nobody quoted, so the catalogue keeps today's and the row's note says so. If the page changes on 1 December, the weekly price-drift check reports it; until a person reads the new figure, GPT-5.6 Sol calls after 30 November may be priced low.

Guarded#

  • A weekly check that the catalogue still agrees with the pricing pages. scripts/price-drift.mjs reads Anthropic's and OpenAI's pricing pages in their Markdown form and compares every published figure with the catalogue: input, output, cache hits, both cache writes, batch, the long-prompt bands and fast mode, 50 models in all, including the "Specialized models" table where the Codex and ChatGPT models are priced, and the Flex, Fast and Ultrafast tables. A figure the page rounds to the cent agrees when it is within half a cent; "$0.40" against $0.375 is drift. A changed price, a model the catalogue lacks, or a page whose tables moved is a finding and fails the scheduled CI run; on pull requests it is printed in the step's log. Freshness says how long ago a provider was read; this says whether its page has changed since. Three of the corrections in 2.7.1 were prices already published for weeks. It never writes a price. GPT-5.5 Cyber and the three Rosalind models are on OpenAI's page and left unpriced on purpose, with the reason in the script: each is behind an approval programme and has no model page to give a context window. Google, DeepSeek, Moonshot, xAI and Mistral publish no page a script reads reliably and stay with a person.