Index methodology
Tokenando Indices
Version 1.10.0. Everything below (formulas, constants, source weights, exclusion rules) renders from the same versioned configuration the settlement job computes with. If a value changes, the version changes and the changelog records it. Already-settled history is not edited except as the revisions policy below allows, and every such case is listed.
ContentsEvery section of the method
Principles
Raw-first: every collector stores the untouched upstream payload before any parsing; settlement reads only stored raw data, never live sources.
Deterministic and idempotent: re-running a settled day must reproduce every value exactly, or the run fails loudly. Settled values are protected at the database level, and the only times they have been changed are listed under “Revisions and estimates policy”.
Drop, do not fudge: a lab or GPU with insufficient coverage is omitted for the day with a recorded reason. No settled value is interpolated, carried forward, or estimated. The 11–29 Oct gap estimates described under TTPI are separate, labelled series and are never stored as settled values.
Everything disclosed: offered prices are labelled offered; volume proxies are named with their bias; every methodology change is versioned and logged.
TTPI: Token Price Index
TTPI(lab, day) = Σ spend(model) ÷ Σ billed_tokens(model) over the lab’s covered models (headline: average price paid) billed_tokens(model) = tokens(model) − BYOK_tokens(model) over the same seven days (from 1.9.0) blended(model) = input_price × 0.75 + output_price × 0.25 TTPI_list(lab, day) = Σ blended(model) × volume_share(model) over the lab’s covered models
Every daily value covers the trailing seven days: the figures OpenRouter reports for each model are totals for the seven days ending on the settled day, so a value for day D describes D and the six days before it, and the series moves more slowly than a one-day figure would. The headline is the average price paid (the realised price): the US dollars buyers paid for a lab’s tokens over the seven days to the settled day, divided by the tokens they bought, per million. OpenRouter stopped reporting spend in dollars after 2026-10-10 and now reports each model’s share of total spend. Since 2026-10-30 we turn shares back into dollars by measuring total spend each day from reference models that bill at list price; at least 8 must agree to within 2%, or the headline drops for the day. Days from 2026-10-11 to 2026-10-29 are not settled, because neither figure was captured. For Anthropic and Google those days carry an estimate, drawn dashed and stored apart from settled values (below); for the other labs they are a gap. Estimates for Google, OpenAI and Anthropic built from a third-party index were published from 1.6.3 to 1.6.5 and withdrawn in 1.7.2. It comes from the spend and token counts OpenRouter reports for each model, so caching discounts and the real mix of input and output tokens are in the number because they are in the bill. Tokens bought with the buyer’s own provider key (BYOK) are not billed at the model price. From methodology 1.9.0 (the value for 2026-11-11 onward) they are left out of the tokens the spend is divided by: OpenRouter’s rankings report none, so each model’s BYOK tokens over the same seven days come from OpenRouter’s per-model daily activity statistics, captured each morning before the settle, and are stored with the value. A model more than 50% BYOK is left out of the value, because OpenRouter’s spend includes its fee on BYOK tokens (measured at 4% to 7% of the price) and that fee would then make up a large part of it. A model whose BYOK count is missing is left out too and counts against the coverage floor below, so a value never mixes the two methods. Values settled before 2026-11-11 count BYOK tokens and are not restated. That pulled them down by roughly each lab’s BYOK share: on 2026-11-10 the change raises Anthropic by 34.2%, OpenAI by 21.0%, Google by 14.7%, xAI by 7.1% and the open-weight labs by 1% to 10%, so the lines step up on 2026-11-11 for this reason alone. MiniMax and Tencent, whose tokens were 73% and 95% BYOK on 2026-11-10, are not published while most of their volume is on models more than half BYOK; their earlier values sit far below what paying buyers were charged. The realised series starts on 2026-09-21, the first day that spend was reported; it is not estimated for earlier days.
The 11–29 October estimate (methodology 1.8.1) uses only OpenRouter’s own figures. For each model and day OpenRouter publishes the price buyers actually paid at each host (input after cache discounts, and output), the tokens each host served, and the tokens sent with the buyer’s own key. From these we rebuild a lab’s realised price for each gap day the way the headline is built: the seven days to the day, spend divided by all tokens. For 11–26 October the true answer is also known as one sixteen-day figure, from OpenRouter’s 30-day spend shares captured on 2026-11-10, and the rebuild is scaled to it with one factor per lab (0.966 for Anthropic, 0.985 for Google). We tested it wherever the truth is known. With the factor taken from 11–26 October, it was within 2.0% of the settled value for Anthropic and 0.8% for Google on each of the eleven days from 30 October to 9 November; with the factor taken from those eleven days, it was within 1.0% and 0.3% of the 11–26 October figure. We publish an estimate only when every such test is within 2%, and state ±2%. It cannot be tested before the gap, because OpenRouter’s daily BYOK counts start on 10 October, and for the same reason Anthropic is estimated from 16 October and Google from 14 October. The other ten labs missed the 2% test (their worst test errors ran from 3.0% to 56%), and their gap stays blank.
The list series prices the same models at their posted rates instead, blending input at 75% and output at 25% with no caching, and weights them by volume. It runs from index inception and was TTPI’s only series until methodology 1.5.0. From 1.8.0 it is settled for closed-weight labs only. An open-weight model’s posted price on OpenRouter is whichever host is cheapest that day, so the series followed which host was cheapest, not what the model costs: DeepSeek’s moved 5% or more on 50 of 84 days. DeepSeek, Meta, Mistral, Qwen, Moonshot AI, Z.ai, MiniMax and Tencent therefore have a realised headline only; their list history up to 2026-11-09 stays as settled and is marked as withdrawn from publication. Routed traffic is far more input-heavy than that blend and mostly cached, so the list series runs well above the realised price; the gap between them is what caching and the traffic mix take off the rate card.
Prices come from the OpenRouter models API, captured at 23:45 UTC as the day’s close; prices are converted to USD per million tokens, and per-model price overrides shadow base prices where present. Free variants and zero-priced listings are excluded: TTPI is a paid-token index. From methodology 1.7.0 (the value for 2026-11-10 onward) it is also a text index: a model that returns images, audio or video is left out, by the output types OpenRouter declares. Values settled before that include those models, which for Google raised the realised price by about 12%; they are not restated, so Google’s line steps down on that day for this reason alone.
Volume weights come from the OpenRouter rankings snapshot captured the following morning (05:10/05:40 UTC), accepted when its newest row is dated the settled day. From methodology 1.8.0 every row of it is used. Each row is the seven-day total ending on the settled day, whatever date it carries (the date is only the last day that model had traffic), as checked against OpenRouter’s own daily figures; before 1.8.0 rows dated earlier were discarded. A model that has just left the price feed, whose spend is still in the window, is priced at its last posted price for up to six days. Variants standard and thinking only (free and batch are excluded). Both series use the same rows, models and floors; a model counts as covered for the realised series only when spend was reported for it. This weight source reflects routed traffic, not any lab’s total global volume: direct enterprise API traffic is invisible to every public index. That bias is the honest cost of an open methodology, and it is disclosed here rather than hidden.
A lab publishes only when it has at least 2 models carrying both a price and volume, and when the priced models cover at least 90% of the lab’s included-variant volume. Below either floor, the lab drops for the day with a recorded reason.
| Lab | Slug | OpenRouter prefixes | Closed-weight |
|---|---|---|---|
| Anthropic | anthropic | anthropic | yes |
| OpenAI | openai | openai | yes |
| yes | |||
| xAI | xai | x-ai | yes |
| DeepSeek | deepseek | deepseek | no |
| Meta | meta | meta-llama | no |
| Mistral | mistral | mistralai | no |
| Qwen | qwen | qwen | no |
| Moonshot AI | moonshot | moonshotai | no |
| Z.ai (Zhipu) | zhipu | z-ai | no |
| MiniMax | minimax | minimax | no |
| Tencent | tencent | tencent | no |
TTPI Price Index: price versus mix
product(model) = the model’s product line (TPPI’s line list), or the model on its own if it is in no line
price(product, day) = Σ spend ÷ Σ billed_tokens over the product’s models (average price paid for the line)
link(a → b) = exp Σ ½(s_a + s_b) × ln(price_b ÷ price_a) over products present on both days,
s = the product’s share of the matched products’ spend (Törnqvist)
PriceIndex(lab, b) = PriceIndex(lab, a) × link(a → b), PriceIndex(lab, 2026-09-21) = 100
mix(lab) = (headline_b ÷ headline_a) ÷ (PriceIndex_b ÷ PriceIndex_a)The TTPI headline is the average price paid: the dollars buyers spent on a lab’s tokens divided by the tokens they bought. It falls when prices fall, and it also falls when buyers move to cheaper models with no price changing at all. Between 21 September and 9 November 2026 Anthropic’s average price paid fell 55% while no list price changed: three new models took 87% of its tokens. A price index is meant to separate the two, so from methodology 1.10.0 each lab has a second series beside the headline, the TTPI Price Index, 100 on 2026-09-21.
It follows product lines, not individual models. A line is one product across its generations (Claude Opus 4.6, 5 and 5.5), using the same published line list as TPPI. A line’s price on a day is the average price paid across its generations, so when buyers move from Opus 5 to a cheaper Opus 5.5 the line has become cheaper and the index falls. When buyers move from Opus to Haiku, no line has changed price and the index does not move; that shift shows only in the headline. Each day is linked to the previous one with a Törnqvist formula: the change in each line’s price, weighted by that line’s share of spend on the two days. The daily links are multiplied into a chain.
How products enter and leave. A new generation of an existing line counts from its first day, through the line’s average. A model that starts a new line, or belongs to no line, has no effect on its first day and joins the weights from its second. A product leaves after its last day with spend; nothing is carried forward. If the products present on both days hold less than 70% of the lab’s spend on either day, no value is published for that day and the next day is linked from the last published one. The 11 to 29 October gap is crossed the same way, as one link from 10 to 30 October.
Why lines and not models. An index that matches individual models treats every new generation as a new product with no effect when it arrives, then reads the old generation’s drifting price as the lab’s price. Old models are not repriced; they lose their heaviest, most cache-efficient buyers to the new one, so what is paid for them per token rises. Matched by model, OpenAI’s prices rose 36% from 21 September to 9 November while its average price paid fell 34%. Matched by line, the change is −2%. The line-matched index errs the other way: it treats a new generation as the same product at a new price, with no allowance for it being better. No accepted measure of model quality exists to adjust by, so we state the assumption rather than hide it. It understates how far quality-adjusted prices have fallen, which is the conservative direction.
What “price” still contains. A line’s price is what was actually paid, so it moves with caching and the input/output mix as well as with rate cards. No Anthropic model changed its list price between 21 September and 9 November: its 47% fall within lines came from cheaper new generations taking the lines’ traffic and from less being paid per token for the same models, mostly through caching. The weekly per-model activity capture that began on 9 November records cached tokens and will allow that part to be split out.
Reading the two lines together. The page gives one sentence per lab and period: the average price paid moved by X, prices within product lines by Y, and the shift between lines by the rest. The two parts multiply: Anthropic’s 0.450 (a 55% fall) is 0.527 from prices within lines times 0.854 from buyers moving to cheaper lines. Labs with no lines defined (Meta, Mistral and Tencent) are chained model by model, so for them the index is model-matched.
The index is settled each morning with the headline, from the per-model inputs stored with the headline, and published under series price_index. Values from 2026-09-21 up to the day before it was first settled were computed afterwards from the inputs stored with each settled headline. They are stored as a separate series, price_index_backcast, drawn dashed, and marked as back-cast on the page and in the API (backcast: true). They are not restatements: no value of this series was published before 1.10.0.
TPPI: Tier Price Index
member(model) = matches a line in the published tier registry (first match wins) TPPI(tier, day) = Σ spend(model) ÷ Σ billed_tokens(model) over the tier’s members (headline, realised) TPPI_list(tier, day) = Σ blended(model) × volume_share(model) over the tier’s members (closed-weight tiers only)
TPPI is the Tier Price Index: one price per class of model, across labs. The classes are Frontier, Production, Open flagship, Everyday, Fast & Quick and Value, as defined for the Follow the Token workshop on 21 September 2026. TTPI follows each lab across all of its models; TPPI follows each class across labs, so it answers what a frontier or an everyday token costs, wherever it is bought.
Membership is a published list of product lines, each placed in one tier, never a price band. A price band would be circular: a model that raised its price would leave its tier, so a tier could only ever look cheap. A line covers every generation of one product (Claude Opus 4.6, 5 and 5.5), so when traffic moves to a cheaper new generation the tier’s price falls, which is how labs mostly cut prices. The workshop named Claude Fable and GPT Astra (Frontier), Claude Opus and GPT Sol (Production), Kimi (Open flagship), Claude Sonnet and GPT Terra (Everyday), Claude Haiku (Fast & Quick), and DeepSeek Flash and GPT Luna (Value); the other labs’ lines are placed by how each lab positions them. Placements are reviewed monthly and every change is a new methodology version. Not yet placed: Tencent’s Hunyuan models (HY3, HY4), whose positioning and weights licence are not yet clear, and small open models such as Gemma, gpt-oss, Llama and Mistral Nemo; embedding, image, audio and code-only variants are outside the tiers by design.
The headline is the realised price, TTPI’s formula applied to the tier: the dollars buyers paid for the tier’s tokens over the seven days to the settled day divided by the tokens they bought, per million, so caching discounts and the real input/output mix are included. It starts on 2026-09-21, the first day per-model spend was reported. The list series prices the same members at posted rates with TTPI’s 75/25 blend, from index inception, and is published only for tiers whose lines are all closed-weight. An open-weight model’s posted price on OpenRouter is whichever host it routes to that day: between 21 September and 6 October 2026 that moved Kimi’s list price 13% and DeepSeek Flash’s 40% a day, while their realised prices moved 1 to 1.5%.
A tier publishes a series only when at least 2 members from at least 2 labs are covered and they carry at least 90% of the tier’s member volume; otherwise that series drops for the day with a recorded reason. Neither series adjusts for quality: a better generation at the same price shows as no change.
Until methodology 1.6.0, TPPI was the Production Price Index: a chained price index and an average price paid over one curated production tier. Those two series stopped on 2026-10-06; their settled values stay in the database as frozen history.
| Tier | Line | Lab | OpenRouter match | Weights |
|---|---|---|---|---|
| Frontier | Claude Fable | Anthropic | ^anthropic\/claude-.*fable | closed |
| Frontier | GPT Astra | OpenAI | ^openai\/gpt-[\d.]+-astra | closed |
| Production | Claude Opus | Anthropic | ^anthropic\/claude-.*opus | closed |
| Production | GPT Sol | OpenAI | ^openai\/gpt-[\d.]+-sol | closed |
| Production | Gemini Pro | ^google\/gemini-[\d.]+-pro (not image|tts|customtools) | closed | |
| Production | Grok | xAI | ^x-ai\/grok-\d (not fast|mini|code|image|imagine|vision) | closed |
| Production | Qwen Max | Qwen | ^qwen\/qwen[\d.]*-max | closed |
| Open flagship | Kimi | Moonshot AI | ^moonshotai\/kimi-k\d | open |
| Open flagship | DeepSeek flagship | DeepSeek | ^deepseek\/(deepseek-v[\d.]+-pro|deepseek-v3|deepseek-chat-v3|deepseek-r1) | open |
| Open flagship | GLM | Z.ai (Zhipu) | ^z-ai\/glm-[\d.]+ (not flash|air|-\d*v\b|vision) | open |
| Open flagship | MiniMax M | MiniMax | ^minimax\/minimax-m\d | open |
| Everyday | Claude Sonnet | Anthropic | ^anthropic\/claude-.*sonnet | closed |
| Everyday | GPT Terra | OpenAI | ^openai\/gpt-[\d.]+-terra | closed |
| Everyday | Gemini Flash | ^google\/gemini-[\d.]+-flash (not flash-lite|image|tts|live|transcribe) | closed | |
| Fast & Quick | Claude Haiku | Anthropic | ^anthropic\/claude-.*haiku | closed |
| Fast & Quick | GPT mini | OpenAI | ^openai\/gpt-[\d.o]+-mini (not image|audio|codex|realtime|search|transcribe|tts) | closed |
| Value | DeepSeek Flash | DeepSeek | ^deepseek\/deepseek-v[\d.]+-flash (not vision) | open |
| Value | GPT Luna | OpenAI | ^openai\/gpt-[\d.]+-luna | closed |
| Value | GPT nano | OpenAI | ^openai\/gpt-[\d.o]+-nano (not image|audio|codex|realtime|search|transcribe|tts) | closed |
| Value | Gemini Flash-Lite | ^google\/gemini-[\d.]+-flash-lite (not image|tts) | closed | |
| Value | GLM Flash | Z.ai (Zhipu) | ^z-ai\/glm-[\d.]+-flash | open |
| Value | Qwen Flash | Qwen | ^qwen\/qwen[\d.]*-flash | open |
TCPI: Compute Price Index
provider_price(gpu) = median of that provider’s per-GPU offers across the day’s captures TCPI(gpu, day) = weighted median of provider_price over all providers (cleared 2.0 · posted 1.0)
GPU offers are captured four times daily (05:30, 11:30, 17:30, 23:30 UTC) from three disclosed sources: the Vast.ai public ask book (one query per GPU, price-ascending so any truncation at the venue’s 64-offer cap is stable and bites the expensive tail), RunPod’s public GPU pricing (secure and community cloud as two distinct providers), and the Shadeform catalogue, which carries posted rates for 19 underlying clouds, each counted as its own provider, only regions with live availability.
All prices are normalised to USD per single GPU-hour (node prices divided by GPU count; Shadeform’s cent-denominated prices divided by 100, cross-verified against posted dollar rates). Offers outside a per-GPU plausibility band are discarded and counted. Every current source posts asked/on-demand rates, so v1 is honestly an OFFERED-price index; the cleared-source weight is declared now so adding a transaction-based source later is a source addition, not a formula change.
A GPU publishes only with offers from at least 3 distinct providers. Interruptible/spot prices are excluded from the headline. They were published as a separate spot series until 2026-10-07; RunPod, its only source, stopped publishing spot bids on 2026-10-08, so the series has no values after that and is no longer shown. The per-provider median-of-offers followed by a cross-provider weighted median means neither a flood of listings from one venue nor a single outlier provider can move the index.
| GPU key | Plausibility band $/GPU-hr | Vast names | RunPod names | Shadeform type |
|---|---|---|---|---|
| H100 SXM | 0.5 – 15 | H100 SXM | H100 SXM | H100 (sxm) |
| H100 PCIe | 0.5 – 12 | H100 PCIE | H100 PCIe | H100 (pcie) |
| H200 | 0.8 – 20 | H200 | H200 SXM | H200 |
| B200 | 1 – 30 | B200 | B200 | B200 |
| A100 SXM4 | 0.2 – 10 | A100 SXM4 | A100 SXM | A100_80G (sxm) |
| RTX 4090 | 0.05 – 5 | RTX 4090 | RTX 4090 | RTX4090 |
| RTX 5090 | 0.08 – 8 | RTX 5090 | RTX 5090 | RTX5090 |
Day attribution & settlement
All days are UTC. For settled day D: GPU offers are captured four times during D; token prices are captured at 23:45 UTC on D; volume weights are captured at ~05:10 UTC on D+1 (the rankings snapshot upstream attributes to D); settlement runs at 06:00 UTC on D+1. Settled values are written once, at that moment; “Revisions and estimates policy” below says when one can change afterwards.
The published record begins at index inception, 2026-08-15. Settlement refuses to produce a value for any earlier day, so the start of the series is a deliberate, enforced boundary rather than wherever data happened to begin. Values computed before inception, while the methodology was still being written, were removed rather than published; the raw observations behind them were kept.
Revisions and estimates policy
Settled values
Each day’s values are computed the following morning from data stored the day before, and written once. We do not edit a settled value because the market moved, because better data arrived later, or because we improved the method. A change of method applies from a published date forward and is recorded in the changelog with a version number.
Corrections
We will correct a settled value in one case: it was wrong under the method published at the time. That means a fault in our code or a corrupted input, not a method we later thought better of. When we correct a value we publish, on the same day, what changed, by how much, why, and which days were affected. The original value stays in our records. An error smaller than 0.5% of the value is recorded in the changelog and corrected only from the next day forward.
Changes of method
We announce a change of method in the changelog before it takes effect, with the reason and its measured effect on recent values. Where a change makes a series step, the chart marks the step. Where it helps, we publish the earlier history recomputed under the new method as a separate, labelled series. It never replaces the values that were settled.
Estimates
When a source fails and a value cannot be computed, the settled series is left blank for that day. We may publish an estimate for the missing days. An estimate is always drawn differently from settled values, is labelled with its method and its tested error, is stored separately from settled values, and is never used to compute another index. If the test of an estimate shows an error wider than 5%, we publish a range, not a single figure. Estimates can be replaced or withdrawn when better evidence arrives; each replacement is recorded and the earlier figures remain available.
When a source is lost
Each index names its primary source and, where one exists, a second source. If the primary source fails or changes, we (1) store whatever it still provides, (2) leave affected values blank and do not carry the last value forward, (3) say so on the index page within one working day, and (4) resume only from a method that has been tested against days when the true value was known, published as a new version. If no such method exists within 30 days, we say whether the series is suspended or ended.
Who decides
A correction, a change of method, or the publication or withdrawal of an estimate is prepared by the engineer responsible for the index and approved by Tokenando’s founder. Both names and the date are recorded with the change. Questions and challenges to any published value go to hello@tokenando.ai; we answer within five working days and publish any correction that results.
Our record so far
Since the indices began on 2026-08-15 we have changed stored values five times. Each time the immutability protection was lifted for one recorded migration in the public repository (drizzle/migrations), with its reason written in the file.
| Date | Migration | What changed | Why |
|---|---|---|---|
| 2026-08-16 | 0008 | Values for two construction days before the official start (2026-08-14 and 2026-08-15) were removed. 2026-08-15 was later settled again as the first day of the record (1.2.2). | Token prices for both days were captured late, because the end-of-day collector did not exist yet, and the days held values from three methodology versions, so neither could be reproduced. |
| 2026-09-25 | 0010 | An open-weight price series for TPPI, computed from 2026-08-15 to 2026-09-24 in testing, was removed. It had not been deployed or published. | An open-weight model’s posted price is whichever host is cheapest that day, so the series moved 20 to 50% a day on host churn. It measured which host was cheapest, not what a model costs. |
| 2026-10-05 | 0011 | Every TPPI value since inception, public for one day, was removed and recomputed from the stored snapshots under methodology 1.4.0. | The original method chained individual models, so a cheaper new generation launched beside an old one registered as no change. 1.4.0 chains product lines. |
| 2026-10-06 | 0013 | TTPI’s posted-price history was relabelled from “headline” to “list” when the realised price became the headline. No value, day or version on any row changed. | The relabelling kept the posted-price series beside the new realised headline (1.5.0). |
| 2026-11-10 | 0020 | A set of estimates for Google and Moonshot AI covering 2026-10-16 to 2026-10-29, published that morning, was withdrawn and replaced the same day. They were stored as estimates, never as settled values. | Checked against ORNN’s index, the Google estimate ran 9 to 12% low on 16 to 20 October, and Moonshot AI’s could not be checked. |
Only the first and third changed settled values that had been published; the second removed a series that was never published, the fourth changed a label, and the fifth withdrew estimates, which are not settled values. The 1.6.3 to 1.6.5 estimates that replaced the withdrawn ones were themselves withdrawn in 1.7.2; no settled value was touched, and they remain in our records. The changelog below describes the methodology changes around each of these.
Sources
| Source | Type | Feeds | Notes |
|---|---|---|---|
| OpenRouter models API | posted prices | TTPI, TPPI prices | official public API; per-model USD-per-token |
| OpenRouter rankings | venue-reported volume | TTPI, TPPI weights; TTPI realised spend | public endpoint; spend in dollars reported 2026-09-21 to 2026-10-10, as shares of total spend since 2026-10-30; routed-traffic bias disclosed above |
| Vercel AI Gateway leaderboards | venue-reported usage shares | collected daily for validation; not yet an index input | CC BY 4.0, attribution to Vercel; token, spend and request shares per lab and model; spend estimated at list prices |
| Vercel AI Gateway models | posted prices | collected daily for validation; not yet an index input | official list prices incl. cache, fast mode and EU regional prices |
| OpenRouter per-host endpoints | posted prices, per host | collected daily for validation; not yet an index input | every host’s price for the 60 highest-volume index models |
| Hugging Face Inference Providers | posted prices, per host | collected daily for validation; not yet an index input | open-weight host prices (Together, Fireworks, Baseten, DeepInfra and others) |
| AWS Bedrock price list | official regional prices | collected daily for a future regional series | US, EU, UAE, Bahrain, Israel; in-region, global, flex, priority and batch tiers |
| Azure retail prices | official regional prices | collected daily for a future regional series | Foundry Models in US, EU, UAE, Qatar; global, data-zone and regional deployments |
| LiteLLM, models.dev | community price databases | quality checks only; never an index input | flag stale or mis-parsed prices before settle |
| Vast.ai ask book | posted (marketplace asks) | TCPI | public search API; ask book only, no utilisation signal |
| RunPod pricing | posted | TCPI | public GraphQL; secure + community as two providers; spot bids stopped 2026-10-08 |
| Shadeform catalog | posted (19 clouds) | TCPI | public API; per-region availability gating; cents→USD verified |
Free JSON API
The headline series are free, keyless, and CORS-open. Responses carry the methodology version and coverage metadata on every value.
| Endpoint | Returns |
|---|---|
| /api/v1/catalog | Indices, entities, date ranges, version, disclaimer |
| /api/v1/ttpi?lab=&series=headline|list|price_index&from=&to= | Token price index, per lab per day ($/MTok); headline is the average price paid, list is posted prices, price_index is the TTPI Price Index (21 Sep 2026 = 100; back-cast values carry backcast: true) |
| /api/v1/tppi?tier=&series=headline|list&from=&to= | Tier price index, per model class per day ($/MTok); headline is realised, list is posted prices |
| /api/v1/tcpi?gpu=&series=headline|spot | Compute price index, per GPU per day ($/GPU-hr); spot ends 2026-10-07 |
| /api/v1/labs · /api/v1/gpus | Registries (config-derived) |
Changelog
| Version | Date | Changes |
|---|---|---|
| 1.10.0 | 2026-11-11 | New series: the TTPI Price Index, one per lab, 100 on 2026-09-21. The headline is unchanged and is now named for what it is, the average price paid. It moves when prices move and also when buyers shift between cheaper and dearer models; the price index moves only when prices move. It chains product lines (TPPI’s published list) with a daily Törnqvist link, so a cheaper new generation of a line counts as a price cut and traffic moving between lines does not count at all. From 21 September to 9 November 2026: Anthropic’s average price paid fell 55.0% and its price index 47.3%; OpenAI −34.3% and −2.2%; Google +13.4% and +16.9%; xAI +9.5% and +9.4%. The difference is buyers moving between lines. No list price of the matched Anthropic models changed in the period; the price part is what was actually paid per line, which includes caching. A new generation of a line counts from its first day; a new line, or a model in no line, joins from its second day; a product leaves after its last day with spend. If the products on both days hold less than 70% of the lab’s spend on either day, no value is published that day and the next day links from the last published one. Meta, Mistral and Tencent have no lines defined, so their index is matched model by model. Settled daily from the first day settled under this version, from the per-model inputs already stored with the headline. Values from 2026-09-21 up to then are a back-cast computed from those stored inputs: a separate series (price_index_backcast), drawn dashed and marked back-cast on the page and in the API. No value of the series was published before, so this is not a restatement, and no settled value changes. API: /api/v1/ttpi accepts series=price_index (unit index; back-cast values carry backcast: true). |
| 1.9.0 | 2026-11-11 | The realised TTPI headline now leaves out tokens bought with the buyer’s own provider key (BYOK). OpenRouter does not charge the model price for them, so they were in the tokens the spend is divided by but not in the spend, which pulled every lab’s price down by roughly its BYOK share. OpenRouter’s rankings report no BYOK tokens; we now take each model’s daily count from OpenRouter’s per-model activity statistics, captured every morning before the settle, and subtract it over the same seven days, model by model. The BYOK tokens removed are stored with each value. Effect, measured on 10 November, the last value settled before the change: Anthropic +34.2% (27% of its tokens were BYOK), OpenAI +21.0% (17%), Google +14.7% (13%), Qwen +9.8%, xAI +7.1%, Z.ai +6.8%, Mistral +6.7%, Moonshot AI +6.5%, DeepSeek +5.0%, Meta +1.3%. The step on 11 November should be about this size. Replayed over 30 October to 10 November the change ranged from +28.5% to +34.2% for Anthropic and +12.0% to +21.0% for OpenAI, because the BYOK share moved within the fortnight; about a third of Anthropic’s fall over those days (−9.8% settled, −6.3% with BYOK out) was its BYOK share rising, not its price falling. The new figure measures what buyers paying OpenRouter were charged. We compared each model’s realised price with OpenRouter’s own effective price for it (what buyers paid at each host after cache discounts) on every settled day from 30 October to 9 November. For models up to 30% BYOK, which carry nearly all of the closed labs’ volume, the median ratio is now 1.00 to 1.04, against 0.77 to 0.99 before; for models 30% to 50% BYOK it is 0.99 weighted by spend (1.11 by median). OpenRouter’s spend still includes its fee on BYOK tokens, which we measure at 4% to 7% of the model price. When most of a model’s tokens are BYOK, that fee is a large part of what is divided by the few billed tokens, so a model more than 50% BYOK is left out of the value, like a model with no spend. MiniMax (73% BYOK on 10 November) and Tencent (95%) are carried almost entirely by such models, so their headlines are not published while that lasts (reason byok_share_above_max). Their values settled until 10 November divided spend by tokens most of which were never billed, so they sit far below what paying buyers were charged: for Tencent’s main models about a tenth of OpenRouter’s effective price. If the BYOK count is missing for a model (the morning capture failed, or its seven-day tokens do not match the rankings), that model is left out of the value and counts against the 90% coverage floor; a lab that falls below it is not published for the day (reason byok_counts_missing). A value never mixes models with BYOK removed and models with BYOK counted. If the whole capture fails, every realised headline drops for the day, and the day can be settled once the capture is taken, because nothing was published for it. Applies from the value for 2026-11-11. Earlier values are not restated: they count BYOK tokens, and the lines step up on 11 November because of this change, not a price rise. The list series, TCPI, TIS and Frontier Premium are unchanged. TPPI’s realised tiers still count BYOK tokens; that will follow as its own version. |
| 1.8.1 | 2026-11-10 | TTPI estimates for the October gap are back for Anthropic (16 to 29 October) and Google (14 to 29 October), drawn as dashed lines and stored apart from settled values. They are rebuilt from OpenRouter’s own per-model statistics: the price buyers actually paid at each host each day, the tokens each host served, and the tokens bought with the buyer’s own key. No third party’s index is used. For 11 to 26 October the true realised price is known as one sixteen-day figure, from OpenRouter’s 30-day spend shares captured on 2026-11-10. The rebuild is scaled to it with one factor per lab: 0.966 for Anthropic, 0.985 for Google. Tested where the truth is known: with the factor from 11 to 26 October, the estimate was within 2.0% (Anthropic) and 0.8% (Google) of the settled value on every day from 30 October to 9 November; with the factor from those days, it was within 1.0% and 0.3% of the 11 to 26 October figure. We publish only when every test is within 2%, and state ±2%. It could not be tested before the gap: OpenRouter’s daily counts of tokens bought with the buyer’s own key start on 10 October. For the same reason, earlier days in the gap are left blank. OpenAI, xAI, DeepSeek, Meta, Mistral, Qwen, Moonshot AI, Z.ai, MiniMax and Tencent missed the 2% test (worst errors 3.0% to 56%) and stay blank for those days. Correction to the methodology text: it said tokens bought with the buyer’s own key are excluded from the realised price. OpenRouter’s rankings report none, so they have always been counted in the tokens spend is divided by. The text now says so. No settled value changes. |
| 1.8.0 | 2026-11-10 | Every row of the rankings snapshot is now used. Each row is the seven-day total ending on the settled day, whatever date it carries (the date is only the last day that model had traffic), as checked against OpenRouter’s own daily figures on all 10,222 rows from 2026-10-20 to 2026-11-08. Until now rows dated before the settled day were discarded; that was a small amount of volume (0.008% of spend on average) but it could remove a spend reference model on the day. A model that has just left the price feed is priced at its last posted price for up to six days, because its spend is still in the window. Every daily TTPI and TPPI value covers the trailing seven days to the settled day; the page now says so. The list series is no longer settled for open-weight labs (DeepSeek, Meta, Mistral, Qwen, Moonshot AI, Z.ai, MiniMax and Tencent), as TPPI already did for open-weight tiers. An open-weight model’s posted price is whichever host is cheapest that day: DeepSeek’s list series moved 5% or more on 50 of 84 days against a largest daily move of 5% in its realised price. Their realised headlines are unchanged and settled history stays as settled; the last open-weight list value is for 2026-11-09. TIS’s like-for-like pairings, which read the list series of an open-weight lab, settle nothing from 2026-11-10 (TIS is not published). New: a revisions and estimates policy, with the public list of the five times stored values have been changed since 2026-08-15. It replaces the earlier statements that settled values “never change” and “are never restated”, which were not accurate. Reproducibility: each settle run now records the git commit it ran from, the settle code tag follows the methodology version automatically (it had stayed at settle@1.6.0 through 1.6.1 to 1.6.5 and at 1.7.1 on 1.7.2), and /api/v1/ttpi and /api/v1/tppi accept inputs=1 to return the per-model inputs stored with each value. Applies from the value for 2026-11-10. Earlier values are not restated. Headline effect of the rankings change on the 30 days tested: at most 0.02%. |
| 1.7.2 | 2026-11-10 | The TTPI gap estimates for 2026-10-11 to 2026-10-29 (Google and OpenAI as dashed lines, Anthropic as a range) are withdrawn. They were scaled from ORNN’s OTPI, and the index should not rest on a third party’s index. Those days show as a gap again. The withdrawn figures are kept in the record, not deleted, and were never settled values. We are working on recovering those days from OpenRouter’s own records. Any figure that comes of it will be published as a labelled, separate series with its method and its error on days where the true value is known. |
| 1.7.1 | 2026-11-10 | Total spend, which turns OpenRouter’s spend shares back into dollars for every realised TTPI headline and TPPI tier, is now read from 29 reference models instead of 11: every model whose billed spend stayed within 2% of its list price on all 20 days from 2026-09-21 to 2026-10-10 when both were reported. On 2026-11-09, 7 of the 11 agreed; 19 of the 29 do. The centre that readings must agree with is now weighted by each model’s spend share, and at least 8 must agree (was 5). Tested on the 20 days with known dollars, the wider pool’s total was within 0.31% of the true total on every day, against 0.54% before. Applies from the value for 2026-11-10. Earlier values are not restated; the change moves headlines by well under 1%. |
| 1.7.0 | 2026-11-10 | TTPI now covers text models only. A model that returns images, audio or video is left out of both the realised and the list series, using the output types OpenRouter declares for each model. Its “tokens” are not text tokens, so its price per million is not comparable with a text model’s. This changes Google the most. Over the seven days to 2026-11-09, nine image and video models were 0.26% of Google’s tokens and 11.8% of its spend, at $10 to $67 per million, and lifted the realised price from about $0.43 to the published $0.484. OpenAI’s image and audio models were under 1% of its spend; no other lab has any. The change applies from the day settled on 2026-11-11 (the value for 2026-11-10). Earlier values are not restated, so Google’s line steps down by roughly 12% on that day. That step is the scope change, not a price cut. Comparing Google across it compares two different baskets. The models left out are recorded with each settled value, next to the models that went in. |
| 1.6.5 | 2026-11-10 | TTPI shows Anthropic for 2026-10-11 to 2026-10-29 as a shaded range, not a line. No single estimate held for Anthropic: its ratio to ORNN’s OTPI moved from 0.835 on 2026-10-10 to 0.764 on 2026-10-30, and the estimate from OpenRouter token counts sat 6% to 8% below the ORNN-based one on every day. Each day’s range is ORNN’s Anthropic figure averaged over the same seven days as our realised price, times every ratio between those two settled values, widened to include the OpenRouter-based estimate where it falls outside. It is 9% to 13% wide. Hovering shows the range, not a single price; the API serves the bounds as series estimate_low and estimate_high and the midpoint as series estimate. |
| 1.6.4 | 2026-11-10 | TTPI adds a gap estimate for OpenAI, 2026-10-11 to 2026-10-29, made the same way as Google’s: ORNN’s daily OpenAI figure averaged over the same seven days as our realised price, times the median ratio on the settled days either side of the gap (0.791). On its own this method is less certain for OpenAI than for Google: fitted on one side of the gap it missed the other side by between 5% below and 12% above, because ORNN’s daily OpenAI figure moves far more than ours. It is published because a second method with no shared inputs agrees with it: OpenRouter’s per-model token counts priced at each day’s list prices came within 0.5% below to 4.5% above it on every day from 2026-10-16 to 2026-10-29. Treat these days as accurate to roughly 5%. The same cross-check failed for DeepSeek (the two methods differed by up to 24%) and for Anthropic (6% to 8% apart on every day), so both stay blank. |
| 1.6.3 | 2026-11-10 | The TTPI gap estimate now covers Google only, for every day from 2026-10-11 to 2026-10-29, and comes from ORNN’s OTPI rather than OpenRouter activity data: ORNN’s daily Google figure averaged over the same seven days as our realised price, times one ratio, the median of our settled Google headline over that average on the settled days either side of the gap. Fitted on one side of the gap, it came within 2.5% of the settled headline on the other. Because those days are ORNN’s figure rescaled, they say nothing independent about ORNN: a comparison of TTPI with ORNN over 2026-10-11 to 2026-10-29 compares ORNN with itself. The 1.6.2 estimates are withdrawn. Checked against ORNN, the Google estimate ran 9% to 12% low on 16 to 20 October; Moonshot AI’s could not be checked, because ORNN does not cover Moonshot. Both were stored as estimates, never as settled values; the withdrawn figures stay in the public repository. OpenAI and DeepSeek stay blank: across the gap the same method missed by up to 12% and 14%, because ORNN’s daily figures for them swing far more than ours. Anthropic stays blank because its ratio to ORNN moved from 0.81 to 0.73 across the gap. ORNN does not cover the other labs. |
| 1.6.2 | 2026-11-10 | TTPI now shows an estimate for Google and Moonshot AI from 2026-10-16 to 2026-10-29, drawn dashed and labelled as estimated. It is stored as a separate series and is never part of the settled headline. The estimate prices each model’s daily prompt, cached and output tokens from OpenRouter’s per-model activity data at that day’s list prices, then applies one factor per lab: the median ratio of the settled headline to the same calculation over 2026-10-30 to 2026-11-08. Tested a week ahead of the days it was fitted on, it came within about 2% of the settled headline for both labs. Other labs stay blank: the same test missed by 3% to 16% for them, about the size of a normal weekly move. Days from 2026-10-11 to 2026-10-15 stay blank for every lab, because their seven-day windows begin before the earliest activity data (2026-10-10). |
| 1.6.1 | 2026-10-31 | OpenRouter stopped reporting spend in dollars after 2026-10-10 and now reports each model’s share of total spend. TTPI and TPPI headlines resume on 2026-10-30 from those shares. Shares become dollars through one figure, total spend, measured each day from a published list of reference models whose billed spend matched list price to within about 1% on every day from 2026-09-22 to 2026-10-10. Tested against those days, when dollars were still reported, the measured total came within about 0.2% of the true total. At least five reference models must agree to within 2%. If they do not, the headlines drop for the day rather than publish an uncertain figure. Days from 2026-10-11 to 2026-10-29 stay blank. Neither dollars nor shares were captured for them, and nothing is interpolated. List series, TCPI, TIS and Frontier Premium are unchanged. |
| 1.6.0 | 2026-10-07 | TPPI is now the Tier Price Index: one price per class of model, across labs, for six tiers: Frontier, Production, Open flagship, Everyday, Fast & Quick and Value, as defined for the Follow the Token workshop on 21 September 2026. TTPI follows each lab across its models; TPPI now follows each class across labs. Each tier is a published list of product lines, each line covering every generation of one product. The workshop named Claude Fable and GPT Astra (Frontier), Claude Opus and GPT Sol (Production), Kimi (Open flagship), Claude Sonnet and GPT Terra (Everyday), Claude Haiku (Fast & Quick), and DeepSeek Flash and GPT Luna (Value). Google, xAI, Qwen, DeepSeek, Z.ai and MiniMax lines are placed by how each lab positions them; the full list is on this page. Like TTPI since 1.5.0, each tier’s headline is the realized price, spend divided by tokens, from 2026-09-21; a list-price series runs from index inception for tiers made entirely of closed-weight models. Open-weight tiers have no list series, because an open model’s posted price on OpenRouter is whichever host it routes to that day: over 21 September to 6 October that moved Kimi 13% and DeepSeek Flash 40% a day, against 1 to 1.5% for their realized prices. A tier needs members from at least two labs, so it is never one lab’s price. The Production Price Index series (a chained price index and an average price paid over one production tier) stopped on 2026-10-06. Their settled values were not changed or deleted; they remain in the database as frozen history and are no longer shown or served. Tier history was computed from the stored daily snapshots. TTPI, TCPI and TIS are unchanged. |
| 1.5.0 | 2026-10-06 | TTPI’s headline is now the realized price: what buyers actually paid per million tokens, from the USD spend and token counts OpenRouter reports for each model each day. Until now TTPI priced tokens at posted list rates with a fixed 75% input, 25% output blend and no caching. Real routed traffic is about 97% input tokens, and most of that input is served from cache at a tenth of the input price or less, so the list figure ran two to five times above what was paid. On 2026-10-04 Anthropic’s list figure was $7.60 per million tokens; its realized price was $1.09. The realized series starts on 2026-09-21, the first day OpenRouter reported spend per model. It is not estimated for earlier days. It covers the same models as before (those with a posted token price), uses the same volume rows, variants and coverage floors, and excludes tokens bought with the buyer’s own provider key (BYOK), which OpenRouter does not bill at the model price. The list-price series continues unchanged, published beside the headline as the list series from index inception (2026-08-15). Its settled history was relabelled from headline to list in migration 0013; no value, day or version on any row changed. Reading the two together shows how much caching and the input-heavy mix take off the rate card. TIS and Frontier Premium are computed from the list series, as before, so they are unchanged. TCPI and TPPI are unchanged. |
| 1.4.0 | 2026-10-05 | TPPI price index restated. It now chains product lines (families) rather than individual models: a line’s price is the usage-weighted price of its generations, so traffic moving to a cheaper new generation of the same line counts as a price cut. Under 1.3.0 a new generation entered at its launch price as no change, so when GPT-6 Luna launched at $0.20 beside GPT-5.6 Luna at $0.45 the index did not move, and over time it could only register price rises on old models. Restated, the index stands at 109.8 on 2026-10-03 where 1.3.0 showed 157.6. Two families that combined separate products were split, so the index never reads a move between products as a price change: Gemini Flash and Gemini Flash-Lite, GPT mini and GPT nano. This is a restatement of published values, which the methodology otherwise does not do; see the revisions policy on this page. TPPI had been public for one day under 1.3.0; every TPPI value since inception (2026-08-15) was removed and recomputed from the stored daily snapshots under 1.4.0, recorded in migration 0011. The average price paid is unchanged in value and was re-settled with it, so both series carry one version. TTPI, TCPI, TIS and Frontier Premium are unchanged. |
| 1.3.0 | 2026-09-25 | New index: TPPI, the Tokenando Production Price Index. It tracks the production tier (Sonnet, Haiku, Luna, Flash, Gemma, Llama, Qwen Flash and similar) across labs, where TTPI tracks each lab across all of its tiers. Frontier flagships have converged on the same posted price; the production tier is where labs compete on price. Membership is a curated list of named tiers with a published rule, reviewed monthly, never a price band: a band would drop any model that raised its price, so the index could only ever look cheap. Open-weight families are included. Labs that sell a single tier are out. TPPI publishes two series. The first is a chained price index (base 100) over closed-weight members that moves only when those models change price. Open-weight members are left out of it, because their OpenRouter price is the cheapest host of the day and flips daily: combined, that host churn moved the index ±15% a day, and an open-only index moved 20-50% a day. An open-weight index is withheld until open models are priced across all their hosts. The second series is the volume-weighted average price paid over all members, the TTPI formula, which also moves when traffic shifts between cheaper and dearer models. Reading them together separates price cuts from adoption of cheaper models. Existing indices are unchanged: TTPI, TCPI, TIS and Frontier Premium compute exactly as under 1.2.2. TPPI history was computed from the stored daily snapshots back to index inception. |
| 1.2.2 | 2026-08-16 | Correction to the 1.2.1 note below: it said the two construction days had token prices captured "roughly a day late". That was accurate for 2026-08-14 (22h57m late) but overstated for 2026-08-15, whose prices were 7h25m late. Posted token prices change on announcement, roughly monthly, so a seven-hour offset does not move the index materially. Index inception accordingly moved from 2026-08-17 to 2026-08-15, and that day settled. It is the first day with a complete input set: GPU offers captured within the day itself, and volume weights captured the following morning exactly as designed. The late price capture is disclosed on the snapshot record rather than smoothed over. 2026-08-14 remains excluded permanently: its prices were nearly a full day late and no GPU offers were collected at all, so two of the three indices could never settle for it. |
| 1.2.1 | 2026-08-16 | Index inception set to 2026-08-17: settlement now refuses any earlier day outright, so the published record has a defined and enforced start. Values settled during construction on 2026-08-14 and 2026-08-15 were removed (migration 0008). Their token prices were captured roughly a day late, because the end-of-day collector did not exist when those days closed, so the snapshots held the following day’s prices, and the methodology changed three times across them, leaving values the current configuration cannot reproduce. Publishing two days that fail our own reproducibility test, at the exact point a reader looks hardest, was not worth two days of history. Every raw snapshot from those days was kept, immutable and untouched, including volume observations that can never be re-fetched from upstream. Nothing irreplaceable was discarded, and the settle-run log of what happened during construction was kept as an audit trail. |
| 1.2.0 | 2026-08-16 | TIS expanded from one open model on one GPU to five models across nine model-and-GPU combinations (DeepSeek V4 Pro and R1, Kimi K2.6, MiniMax M3 and gpt-oss 120B on H100 SXM, H200 and B200), each at three serving speeds, all sourced from InferenceX comparison pages with per-pairing citations. Readers choose the combination rather than inheriting ours. Frontier Premium is now restricted to frontier-class open models. gpt-oss 120B is excluded from it: at roughly 60M tokens per GPU-hour it would set the cheapest self-hosted cost by being small and fast, turning the headline ratio into a statement about model size rather than hosting economics. It remains fully available in TIS. Pairings that answer no real question are no longer settled: an open-weight lab’s API against somebody else’s open model. Only like-for-like (the lab that serves this exact model) and substitution (closed labs, which cannot be self-hosted at all) are published. Reproducibility checks are now scoped to the methodology version that produced a value. A value settled under an earlier version is frozen history; it cannot be re-derived from today’s constants and is never overwritten or reported as a mismatch. Without this, the first constant change would have made every historical day fail to re-settle, contradicting the purpose of versioning. |
| 1.1.1 | 2026-08-16 | Throughput rows now record which lab publishes the open weights, so the index can distinguish a like-for-like pairing (a lab’s own API vs self-hosting its own model, a clean hosting-margin measurement) from a substitution pairing (a closed lab’s API vs switching to a different, open model, which trades away model quality too). The two answer different questions and are now reported separately instead of in one undifferentiated list. Documented that the implied self-hosting cost assumes full utilisation of the rented GPU and excludes idle capacity, failed requests and operations overhead. It is a floor, not a forecast. Labelling and presentation only: no computed value changes, and no settled value was altered. |
| 1.1.0 | 2026-08-16 | TIS now settles three serving operating points per (model, GPU), namely volume, balanced and latency, instead of one. Published benchmark curves span roughly 11x between the batch and low-latency ends, which moves the Frontier Premium by more than the difference between any two labs; picking one point silently was indefensible, so all three publish and the reader can switch. The "balanced" point (interactive serving, ~116 tokens/sec per user) is the documented headline: the API default, the citable figure, and the number in the daily summary. Throughput table reseeded from InferenceX (SemiAnalysis) for DeepSeek V4 Pro 1.6T on B200 at FP4, replacing the unverified DeepSeek V3 / H200 placeholder. InferenceX reports H200 as unmeasured for this model. Operating points are read from an interpolated published curve, disclosed as such. Weekly staleness check added: verified throughput citations older than 90 days, or whose source URL stops resolving, raise an alert rather than ageing silently. |
| 1.0.0 | 2026-08-16 | Initial methodology: TTPI (volume-weighted blended $/MTok per lab, 0.75/0.25 input/output share), TCPI (weighted median $/GPU-hr per GPU from disclosed posted sources), TIS and Frontier Premium (verified-throughput rows only). Sources: OpenRouter models API (prices), OpenRouter rankings snapshot (volume weights; disclosed with bias), Vast.ai ask book, RunPod GraphQL, Shadeform catalogue (19 clouds). Drop rules: labs below 90% volume coverage or under 2 priced models drop for the day; GPUs under 3 distinct providers drop for the day. Missing data is never interpolated or carried forward. |
How prices for individual models are sourced and verified is documented separately on the About page.
Tokenando indices are informational reference data, not financial products, price quotes, or investment advice. Offered prices are labelled offered; sources, weights, and exclusion rules are fully disclosed in the methodology. Settled values are written once; a correction is published with its reason and listed in the revisions policy, never made silently.
