Method

One pure function over a usage row turns provider tokens into dTok, so a price per million dTok is comparable across providers and mixes. Cache reads count at a tenth of an input token in the unit of account; a delivered-token diagnostic counts them at full weight, and the ratio of the two is the cache efficiency.

Base weights by billing class; output classes take the grade's output ratio.

input_uncached         1.00
input_cache_read       0.10
input_cache_write_5m   1.25
input_cache_write_1h   2.00
embedding_input        1.00
tool_call_metered      1.00
output_visible / output_reasoning  G1  5.00
output_visible / output_reasoning  G2  5.00
output_visible / output_reasoning  G3  4.00
output_visible / output_reasoning  G4  3.00
long_context (input classes)  ×2.00;  batch  ×0.50

Sanity check: a model priced at a 5:1 output-to-input ratio yields the same $/M dTok whatever its input/output mix. A price per dTok that swings with mix on such a model is a defect, not a finding.

What it simplifies, and the direction of each bias

1. Grade comes from a static map of model slugs, not from measurement.A grade dispute is a configuration change, not a data problem.None, but it is an opinion.
2. No reference-tokenizer recount; the reference quantity is empty.No substitution detection and no comparability across tokenizers.None within one provider.
3. Reasoning tokens are folded into visible output where the source does not separate them.The reasoning cap of 4.0 can never bind.Overstates visible output; understates nothing.
4. Cache-write duration defaults to the five-minute weight (1.25) where the provider does not expose it.Exact where itemised; assumed elsewhere.Understates dTok on one-hour cache workloads by up to 60% of that class.
5. Context tier is always short: daily aggregates cannot resolve per-request context.The ×2 long-context multiplier never applies.Understates dTok on long-context traffic, and therefore overstates $/M dTok.
6. Image, audio and container-hour lines carry no dTok.Their cost counts; their dTok does not.Overstates $/M dTok on multimodal workloads.

A closed-form normal approximation to cumulative consumption over the rest of the term: no simulation, no seed, deterministic. The headline outputs are the expected value left unused at expiry and the lock date — the first date after which increased consumption can no longer recover it. The probability of stranding is computed and not headlined, because it saturates.

L = 5.0   Z90 = 1.2815515681605769
mu = mean(daily, last 28d);  sigma = sd(daily, last 56d)
g28 = clamp(mean(last 28d) / mean(prior 28d), 0.5, 2.0);  gd = g28^(1/28)
cum = sum(mu * gd^t, t < H) + step_changes;  sd = sigma * sqrt(H * L)
p10, p50, p90 = cum - Z90*sd, cum, cum + Z90*sd
expected_stranded = max(0, R - p50);  p_strand = Phi((R - cum) / sd)
strand_lock_date = first d with p_strand > 0.95  (binary search)

Below the minimum history the forecast refuses: it prints the observed and required run rates side by side and no dollar figure.

It is unbacktested. It has not been graded against realised outcomes, and every report that prints it says so in that word.

1. Gaussian tails on a right-skewed quantityDaily spend is roughly log-normal; the normal approximation understates upside consumption and so overstates stranding on fast-growing accounts. The bias flatters the pitch, which is why it is printed.
2. L = 5 is a hand-set constantIt stands in for the autocorrelation of a weekly cycle. First replacement, once there are at least 3 tenants with 120 days each: estimate it per tenant as L = 1 + 2·Σ(k=1..14) ρ_k, over 14 lags.
3. The probability of stranding saturatesThe gap between what remains and expected consumption grows with the horizon while the uncertainty grows only with its square root, so the probability sits near zero or one for most commitments. Arithmetic, not a defect — and the reason it is not the headline.
4. No weekly seasonalityHarmless for horizons of 28 days or more; material below about 21 days, where the forecast is withheld anyway.
5. No jump processA launch or a price move is not modelled. The substitute is a step change entered at intake, which covers the moves that are actually known.
6. No cold-start poolingBelow 28 days of history the model refuses and prints run rates only.
7. Attribution is assumedOne commitment per usage source; the order of drawdown across several commitments is not yet modelled.
8. It is unbacktestedSaid on the page, in that word.

Five outputs from the customer's own usage and published rate cards: the realised committed rate; the same trailing mix replayed at today's list; the cheapest equivalent grade at list across providers; the units still owed; and the effect of announced list-price moves on the customer's own mix, split by committed and on-demand share.

List-price comparison, not a market index. While posted or modelled rates make up more than 80% of the basis, the spread in dollars is withheld and printed greyed beside its literal basis mix.

How posted rates are replaced: a provider-consented, countersigned desk deal first; attested order-form rates captured at intake next; anonymised panel medians once enough tenants contribute; venue quotes and prints last.

The spread is a rate comparison. It is not an accounting conclusion and does not say what a customer must book.

Every slice has a ceiling and a floor before any print exists. The buyer will not pay more, all in, than it would pay the provider at list for equivalent units; the seller will not accept less, net of fees, than its best alternative. Prices are signed: where the committed rate is above list the ceiling can be negative, and the seller pays the buyer to take the obligation on.

ceiling = list − provider_billing − buyer_fee
floor   = seller_fee + best_alternative
status  = empty  when floor > ceiling  (refused; no waiver override)

An empty corridor — floor above ceiling — is refused. The launch fee waiver never overrides it.

While the only comparator is a list-rate equivalent, two red limbs apply: a price below the floor ratio of list, or below a fraction of the seller's own attested rate. The green, amber and red bands against an index apply only once an unflagged series exists.

Amber holds are reviewed within 4 business hours.

Red attempts inside the window: a provenance re-audit, then suspension.

A series point is published only when enough independent transactions stand behind it. Below the threshold it is labelled an assessment, never an index. Related-party rows and rows held for review are excluded before counting.

Points more than this far from the last unflagged point are held for review.

A point may be revised once, within 48 hours of publication.

Below the threshold: assessment.

Candles, bars and the range area are drawn only at Week and Month, because one value is published per day. Each is computed on your screen from the days published in that week or month: it opens at the first published day, closes at the last, and reaches the highest and the lowest. No weekly or monthly value is published.

A week or month with no published day is left out. A single published day is drawn as a flat tick. A candle that holds an assessment day is grey; the others are coloured by direction only. A week or month still in progress is drawn lighter. A candle never spans a methodology change or a suspension: the week or month is split there.

The seller pays a share of face value on success; the buyer pays nothing except where a buyer-facing guarantee exists. Fee rates are founder inputs and are shown as estimates until they are closed.

A fast screen, per commitment, of whether a capacity commitment looks like a lease or a service under IFRS 16 B20. Token, throughput, prepaid and capacity commitments screen as services; dedicated or named hardware needs judgement. On its own the screen never concludes that a commitment is a lease.

Every verdict is a draft until a named person reviews it, and the uncommenced-commitment export reads reviewed rows only. It is not an accounting conclusion.

Every load passes ten checks in order before a row is kept, from the header match to the month-total reconciliation against the stated invoice. A re-load of the same file adds nothing.

The ledger, the payment processor and the provider's own billing observation are compared per counterparty. A break above $0.01 opens a severity-two incident and blocks that counterparty's next settlement.

Charts on this site are drawn with the software below, under its licence.

• Lightweight Charts™

  Created by

  TradingView ↗

  Licence

  Apache-2.0 · lightweight-charts 5.2.1

  Attribution notice

  Copyright 2023 TradingView, Inc.

  This software incorporates several parts of tslib (https://github.com/Microsoft/tslib, (c) Microsoft Corporation) that are covered by BSD Zero Clause License.

• TradingView Lightweight Charts™ — HLC area series plugin

  Created by

  TradingView ↗

  Licence

  Apache-2.0 · @tradingview/lwc-plugin-hlc-area-series 1.0.0

  Attribution notice

  TradingView Lightweight Charts™ — HLC area series plugin

  Copyright (c) 2026 TradingView, Inc. https://www.tradingview.com/