Dimension
What a price is charged by. A sku does not have "a price"; it has a price per dimension, and most skus carry several.
Dimensions are registry rows in witness.price_dimension, each belonging to one component class.
The registered dimensions
Language models (model):
| id | Label | Unit |
|---|---|---|
input_per_mtok | Input price | USD per 1M tokens |
cached_input_per_mtok | Cached input price | USD per 1M tokens |
cache_write_5m_per_mtok | Cache write price, 5 minute retention | USD per 1M tokens |
cache_write_1h_per_mtok | Cache write price, 1 hour retention | USD per 1M tokens |
output_per_mtok | Output price | USD per 1M tokens |
batch_input_per_mtok | Batch input price | USD per 1M tokens |
batch_output_per_mtok | Batch output price | USD per 1M tokens |
priority_input_per_mtok | Priority input price | USD per 1M tokens |
priority_output_per_mtok | Priority output price | USD per 1M tokens |
flex_input_per_mtok | Flex input price | USD per 1M tokens |
flex_output_per_mtok | Flex output price | USD per 1M tokens |
output_per_image | Output price, per image | USD per image |
batch_output_per_image | Batch output price, per image | USD per image |
A model that emits images is billed by the image rather than by the token, so the class carries two units rather than one. That is why the unit travels on the figure and is not a property of the class.
Agent tools (tooling):
| id | Label | Unit |
|---|---|---|
tool_call_per_1k | Per 1,000 calls | USD per 1,000 calls |
page_fetch_per_1k | Per 1,000 pages | USD per 1,000 pages |
container_hour | Container hour | USD per container-hour |
session_hour | Session hour | USD per session-hour |
storage_per_gb_day | Storage | USD per GB-day |
A fetched page is its own meter rather than a kind of call. One call can return one page or several, and three of the four sellers on page_fetch_per_1k price the page rather than the call. Putting both on one scale would invite a reader to divide one by the other.
session_hour is a hosted session billed for as long as it runs, and neither seller on it uses that name: one sells a managed agent session and the other sells a browser hour. The id names the meter, as input_per_mtok does; sharing one is not a claim that two products are the same, which is what a sku's product slug says and these two do not share.
Embedding and rerank (retrieval), both USD per 1M tokens:
| id | Label |
|---|---|
embed_per_mtok | Embedding price |
rerank_per_mtok | Rerank price |
Document extraction (extraction), both USD per page:
| id | Label |
|---|---|
page_ocr | Text from a page |
page_layout | Structure from a page |
A page is not one thing across these providers. page_ocr is a page read for its characters; page_layout is a page read for its tables, form fields and layout blocks, and what that includes differs by provider. So every extraction sku carries a page attribute saying which kind of page its figure is for, and the boards print it under the rate. Two per-page figures compared without it are two different products under one heading.
Speech recognition and synthesis (speech):
| id | Label | Unit |
|---|---|---|
audio_minute_batch | Pre-recorded transcription | USD per audio-minute |
audio_minute_streaming | Streaming transcription | USD per audio-minute |
tts_per_mchar | Synthesis price | USD per 1M characters |
Transcription is two meters because every seller on the board sells it twice, once for audio that has already been recorded and once for audio arriving live, at two different prices. The two share a unit and are still two meters, which is the same distinction embed_per_mtok and rerank_per_mtok carry: a unit is what a figure is measured in and a meter is what is being sold.
That is also the admissibility test for this class, and it is stricter than it sounds. A seller quoting one recognition rate covering both is quoting neither of these meters, so the record declines it rather than filing the figure under whichever of the two looks closest.
The audio-minute is the audio's own minute, not the wall clock. A pre-recorded service that returns in ten seconds still bills the minute it read.
GPU compute (compute):
| id | Label | Unit |
|---|---|---|
gpu_hour_on_demand | On-demand list price | USD per GPU-hour |
Machine translation (translation), both USD per 1M characters:
| id | Label |
|---|---|
translate_input_per_mchar | Input price |
translate_output_per_mchar | Output price |
Only the pay-as-you-go tier is registered. Every seller on that board also prices a volume commitment, and a price for a promise about next year is not a price a reader can take today.
The two meters are never summed. Three of the five products bill the characters you send and carry no output figure at all; the other two are large language models billing both ends. Google states on its own page that its ten dollar each way product is cost equivalent with its twenty dollar one way product, which holds at a one to one character ratio and at no other, so the board prints the ends a buyer will actually send rather than an equivalence resting on an assumption nobody made. Image generation (image):
| id | Label | Unit |
|---|---|---|
image_per_image | Per image | USD per image |
batch_image_per_image | Batch price, per image | USD per image |
image_per_megapixel | Per megapixel | USD per megapixel |
Two units again, and for a sharper reason than the model class has. Some sellers charge one price for an image whatever size it is; others charge by the area of what comes back. At a one-megapixel output the two can print the same number and still be different products, because at two megapixels one of them has doubled. So image_per_megapixel is a meter of its own rather than the same meter converted, and no spread is ever drawn across the two.
image_per_image is not output_per_image. The latter is the model class's meter for a language model whose output happens to be a picture, priced beside its token rates on a token page; the former is an image product sold on its own. A figure lives under one of them and never under both.
Video generation (video):
| id | Label | Unit |
|---|---|---|
video_per_second | Per second | USD per second |
One meter, because one seller is read so far and it prices one way. What the figure buys is on the sku rather than in the unit, and two things are on it that move a price further than anything else in this class: the resolution, where one model spans $0.05 to $0.60 a second, and whether audio is included, because a rate for video with audio is not a rate for video without it.
The same page states a per-second equivalence for token-billed video and calls it approximate. It is not read. An equivalence the seller itself hedges is not a list price for this unit, and the rule that refuses a from-priced GPU refuses this for the same reason.
Music generation (music):
| id | Label | Unit |
|---|---|---|
music_per_track | Per track | USD per track |
One meter, for the same reason as video: one seller is read so far and it prices one way. The page bills per request and a request returns one track, so the two are the same thing counted from either end.
The length is on the sku rather than in the unit, and nothing divides by it. The same seller charges $0.04 for a thirty second clip and $0.08 for a full song, and the record carries those as two products rather than one rate at two durations. Turning the clip into a price per second would invent a unit the page does not sell: it sells a clip, not thirty seconds. The seller writes "per song" against the clip row too, which is its wording rather than a mistake this record corrects.
A row carries only the dimensions its source publishes
If a provider's page does not quote a batch rate, that sku has no batch_input_per_mtok, and the cell is absent from the tape rather than present and zero. Absence is a fact about the market and gets reported as one.
Sanity bounds are data, not parser code
Each dimension carries sanityMinUsd and sanityMaxUsd. A parsed figure outside them is rejected as a probable misread rather than entering the record.
These live on the registry row rather than inside the adapter so that widening one is a data change with a reason attached, not a code change buried in a parser. Two of them have already been widened for exactly that reason: the model input ceiling was $100 until OpenAI listed o1-pro at $150, which made the bound a limit on Vitna's coverage rather than on what a parser could plausibly have got wrong.
Why every floor is above zero
The floor is not a claim that nothing is free. It is there because zero is what a broken read returns.
A label that matched a navigation menu, a window that stopped before the figures, a cell index that ran past the end of a row: each produces a nothing that arithmetic turns into a nought, and each would otherwise put a free model on a public board with no error raised anywhere.
A page that genuinely publishes $0.00 is a different thing and a real fact about a market. The two are told apart by evidence rather than by declaration: a figure carries the substring the document actually printed, so a zero has to be shown to have been read rather than asserted to have been.
An add-on is its own meter or it is absent
Every speech seller on the board prices diarisation, redaction, keyterm prompting and the rest as separate per-minute rates beside the transcription rate. None of them is registered and none of them is folded into the figure above it, which is what makes the headline rate readable as a headline rate.
The rule generalises past speech: an add-on gets a dimension of its own when the record decides to carry it, and until then the record carries the base rate and says so. Folding one in would produce a number no document prints.
A cache read and a cache write are two meters, and a write is two more
cached_input_per_mtok is the cached read rate: what a request pays for tokens it found already in the cache. Filling the cache is a separate charge, priced above uncached input, and it is the two cache_write_* meters.
Those two are one meter split by how long the seller promises to keep what was written. The same tokens go in either way; a five minute window and an hour are two prices for two promises, printed side by side on the same row of the same page. That is the definition batch_*, flex_* and priority_* already use, so a retention window joins the rule about service tiers below rather than starting a new one.
Why two dimensions rather than one carrying the window as an attribute. Three things decide it and none is taste.
A sku's attrs is a property of the sku, and one model carries both rates at once. Putting the window there means one sku per window, which is the shape the context band refused for the same reason: it doubles every count the boards print and makes a succession field describe a lineage that is really one product.
A spread groups by dimension. One dimension would put a five minute write and an hour's write in one group and print the gap between them as a dispersion. It is not one. It is the distance between two products, and the record's own machinery would have manufactured exactly the misreading this meter has to be safe from.
And the dimension id is what a figure carries everywhere it goes, into the tape, into a bill line, into a citation. With the window in the id there is no reading of $6.25 that could be the hour's rate.
The two names are names, not one name with a number in it. A seller pricing a thirty minute window gets a third dimension rather than a third value of a field, which is not hypothetical: Novita's pricing payload already carries fields for a five minute and a thirty minute write, both empty of figures on 2026-09-10.
One seller is on these meters. Four more publish a write rate and none of them is filed, for reasons worth keeping apart because only one of them is about the market. Read 2026-09-10.
- Anthropic prints both windows in its own pricing table, and is read.
- OpenAI prints a cache-writes column in each of its four tier tables and names no retention window anywhere on the page, so there is no registered meter its figure belongs to. The temptation is the part worth recording: every row prints its write at exactly 1.25x its own input rate, and 1.25x is the five minute multiplier at every seller who publishes both windows. An inference that clean is exactly the kind this record does not make.
- DeepInfra and Novita carry theirs inside a script payload the HTML readers strip before any adapter sees it. A different parse strategy, not a new field.
- Azure lists cache-write meters in the Retail Prices API that the registered query does not ask for, and its meter names say nothing about a window either, so it is held by both of the reasons above at once.
Only OpenAI's is a document that has been read and does not say enough. The other three are documents nobody has asked the right question of yet.
Novita has since been asked. Its pricing page now prints a write column at five minutes and at an hour, and the adapter has read both since 2026-11-02, so two sellers are on these meters and the second is the first in a peer set.
Google is absent from these meters and is not one of those cases. It prices cache storage, per million tokens per hour, which is rent for holding a cache rather than a charge for filling one, and a third unit as well as a third meter. An hour appears in that figure and in cache_write_1h_per_mtok, and the two have nothing to do with each other.
rerank_per_mtok is still the only rerank meter. Some providers sell reranking per search instead. That is not a point on the same scale, so it gets its own dimension when a source publishes one, rather than being converted into this one at an assumed document count.
What a cache write can and cannot be set against
Asked because a meter that draws no board should say so rather than have somebody find out by drawing one.
Against a cache read: never. A write and a read are different events at different prices and the dispersion between them is not a fact about a market. Nothing has to remember this. The spread groups by dimension and they are two.
Against a write at another window: never. Five minutes of retention and an hour of it are two products, and the gap between them is the price of the longer promise rather than a disagreement between sellers. This is guarded the same way and only because the window is in the id.
Against another seller's write at the same window: yes, with a note. Both meters are in the model board's own unit, so nothing bars them from a spread. A dispersion on this board is several sellers of one product, grouped by the product slug an open-weight row carries. Novita's write columns reached three such peer sets in the snapshot of 2026-11-03 (Kimi K3, Qwen3.8 Max and Qwen3.8 Flash), and Novita is the only seller filing a write in each, so no write spread is drawn yet. The reason is coverage rather than category.
The note is the one thing that had to be settled before the first such spread, and it was settled when those peer sets arrived. Sellers cache in blocks, and the block size is theirs to choose: DeepInfra's own page states a per-model granularity running from 1,024 to 8,192 tokens with the remainder billed as standard input. Two sellers quoting the same rate per million tokens at the same retention window are therefore not necessarily selling the same thing, in the way two per-page extraction figures are not. So every answer that sets one seller's write against another's carries it as blockNote: the spread and the peer range a quote check reads. No other meter does.
A character is not a token
translate_input_per_mchar is per million characters. input_per_mtok is per million tokens. The record does not convert between them and will not.
A conversion needs a characters-per-token ratio, and no such constant exists. It is a property of a tokenizer and a language at once: near four characters to the token for English, far denser for Chinese, Japanese, Korean and Thai. Translation crosses languages by definition, so a single ratio would be wrong by a multiple in exactly the work this class is about.
The meters are not counting the same events either. Google states that markup, whitespace and untranslated characters are all charged, and that a request with no content is charged one character. No token meter behaves that way.
So both prices stand in their own unit, side by side, and the ratio between them is a measurement somebody takes on their own corpus. That is the same rule the record already applies to quality: measured on your own documents, or not claimed. A dedicated translator priced per million characters standing beside a general model priced per million tokens is a comparison worth having, and it is worth having in two units rather than in one invented one.
A service tier is a dimension
Providers sell the same model at more than one speed. A batch tier runs the work whenever there is room and costs less; a priority tier runs it sooner and costs more; a flex tier sits between standard and batch. Each is a separate rate the provider publishes on the same page, and each is registered as its own meter.
That is what batch_input_per_mtok has always been, so the six tier meters are one rule rather than three special cases: batch_*, flex_* and priority_* are the same model billed under a different service promise, and a buyer chooses between them per request.
priority_* is named for the meter rather than for what any one seller calls it this year. OpenAI's page prints "Fast pricing data" and states in its own footnote that the tier was called Priority processing until 30 July 2026 and that service_tier: "priority" is still accepted; Anthropic prints "Fast mode pricing" and takes speed: "fast". An id tracking the marketing word would have needed renaming once already, so the id is Vitna's and the locator quotes the page.
Sharing a dimension is not a claim that two sellers' tiers are one product. A dimension is the meter; a peer set is the product, and these two sellers share none, so no spread is drawn across them. container_hour already carries that distinction for OpenAI's sized containers against Anthropic's unsized one.
A context band is not a dimension
Some providers price one model differently by how much context a request carries. That is a property of the price rather than a new meter, so it travels on the figure as a band rather than as a dimension of its own.
The other option was to register a second long-context dimension beside input_per_mtok, and it was refused because it would have meant every consumer learning a second name for the same meter, and every unbanded sku carrying a column it never fills. See the tape.
Money
Amounts are numeric(14,6) USD, in the unit the dimension declares. Six decimal places because a per-token price is quoted per million tokens and dividing back to a single token has to stay exact.
