Methodology · gii-v1 · Full specification

How the print is built.

The full specification is published on this page, versioned under the gii-v1 identifier. Every constant is a stated convention, every print carries a content hash, and a worked one-day example lets anyone recompute the headline number.

What the series measures

The headline series GII-OR-SPEND-ALL measures public list-price notional daily intelligence spend on OpenRouter production traffic (top-50 models, excluding the aggregate other row):

\hat{S}_t \;=\; \sum_{i} V_{i,t}\,\cdot\,\frac{\hat{P}_i}{10^{6}}

USD per UTC day t

$V_{i,t}$ is total_tokens from the public daily rankings. $\hat{P}_i$ is the standardized list price from the print-date tariff snapshot. The result is public notional spend (what observed traffic would cost at posted list prices), not realized invoice data.

Inputs

Public inputs only

Volumes: OpenRouter rankings-daily, the daily top-50 model token totals plus one aggregate other row. Snapshots archived and versioned since 2025-01-01.
Tariffs: OpenRouter models-v1 catalog, the lowest routed public list prices per model, snapshotted daily.
No solicited quotes, no broker feeds, no private data. If a price is only available through sales contact, it is not an input.

Calculation

Standardized list price

\hat{P}_i \;=\; 0.75 \cdot P_i^{\mathrm{input}} \;+\; 0.25 \cdot P_i^{\mathrm{output}}

USD per 1M tokens

The 75/25 blend is a stated convention, not an estimate. Every print also publishes the 90/10 and 50/50 bounds as a blend convention band, so the headline level reads as a banded convention rather than a false-precision point.

Basket construction

\hat{S}^{\,\mathrm{14D}}_{i,t} \;=\; \sum_{\tau = t-13}^{t} V_{i,\tau}\,\cdot\,\frac{\hat{P}_i(t)}{10^{6}}

trailing 14-day spend

Models are sorted by trailing 14-day spend and added until cumulative share reaches 95% of total top-50 trailing spend. The current basket holds 25 models capturing 95.3% of trailing spend.

Derived unit-cost series

I_t \;=\; \sum_{i \,\in\, B_t} W_i \, \hat{P}_i

14D spend weights · USD per 1M tokens

\hat{I}^{\,\mathrm{daily}}_t \;=\; \frac{\hat{S}_t}{\sum_i V_{i,t}} \times 10^{6}

same-day volume weights

GII-OR-UNITCOST-ALL is the basket-weighted unit cost $I_t$ . The daily volume-weighted variant is published as a diagnostic. Neither is the headline; spend is.

Reproducibility

Recompute the print yourself.

Three published artifacts trace the path from raw rankings to the headline number: the eligibility funnel, a worked one-day example, and the blend-convention sensitivity table.

Eligibility funnel · print day 2026-06-07

Stage	Rows	Token share	Note
Raw ranking rows	51	n/a	Includes aggregate other row
Model rows after excluding other	50	100.00%	OpenRouter top-50 model universe
Rows with accepted v1 identifier	50	100.00%	Native permaslug is the v1 join key
Rows with public tariff	49	99.63%	1 model(s) lack a tariff
Coverage-qualified print	49	99.63%	Above 50% validation threshold
95% trailing-spend basket	17	53.58%	Captures 95.14% of trailing spend

Worked one-day example · top 10 rows

#	Model	Tokens	Tariff $/1M	Daily spend	Weight
1	anthropic/claude-4.7-opus-20260416	250,604,609,065	10.00	$2.5M	33.0%
2	anthropic/claude-4.6-sonnet-20260217	219,652,010,422	6.00	$1.3M	19.4%
3	anthropic/claude-4.8-opus-20260528	201,060,276,031	10.00	$2.0M	12.7%
4	openai/gpt-5.5-20260423	39,303,395,876	11.25	$442.2K	9.0%
5	anthropic/claude-4.6-opus-20260205	43,854,970,370	10.00	$438.5K	8.7%
6	google/gemini-3.5-flash-20260519	42,293,718,815	3.38	$142.7K	3.0%
7	openai/gpt-5.4-20260305	19,513,280,563	5.63	$109.8K	2.5%
8	google/gemini-3.1-pro-preview-20260219	25,499,388,305	4.50	$114.7K	2.0%
9	google/gemini-3-flash-preview-20251217	125,880,415,605	1.13	$141.6K	2.0%
10	deepseek/deepseek-v4-pro-20260423	304,435,953,783	0.54	$165.5K	1.4%

The full 17-row audit table ships with the data package.

Blend-convention sensitivity · 2026-06-07

Input/output blend	Notional spend	Unit cost $/1M	Relative to 75/25
90/10	$5.9M	1.203	0.707x
75/25	$8.4M	1.702	1.000x
50/50	$12.5M	2.533	1.489x
25/75	$16.6M	3.365	1.977x

Input integrity

Forensic screens on the source panel

The pipeline applies forensic-accounting conformance tests to the pooled model-day token panel: Benford first-digit, second-digit, and first-two-digit tests with Nigrini conformity thresholds; last-two-digit uniformity; and exact-duplication analysis. Current results: close conformity on all digit tests, no round-number excess, zero duplicates across 26,200 model-days. A restatement audit shows zero restatements across the archived history.

Framing

Screens, not attestation.

These diagnostics screen for signatures of fabricated, rounded, or smoothed source data. They are not provenance authentication, which remains a settlement-grade attestation requirement on the governance roadmap.

Error model

The gap to realized spend.

The difference between realized and notional spend decomposes into five documented terms, listed below with the direction and bound of each where known.

Error model

S_t - \hat{S}_t \;=\; \varepsilon_{\mathrm{blend}} + \varepsilon_{\mathrm{routing}} + \varepsilon_{\mathrm{discount/cache}} + \varepsilon_{\mathrm{coverage}} + \varepsilon_{\mathrm{mapping}}

realized minus notional

$\varepsilon_{\mathrm{blend}}$ : the fixed 75/25 prompt/completion convention (bounded by the published band).
$\varepsilon_{\mathrm{routing}}$ : provider routing and fallback paths below the lowest-routed list price.
$\varepsilon_{\mathrm{discount/cache}}$ : private terms, credits, prompt caching, successful-run billing.
$\varepsilon_{\mathrm{coverage}}$ : top-50 cutoff and the excluded aggregate other row.
$\varepsilon_{\mathrm{mapping}}$ : slug, taxonomy, or tariff-linkage mistakes.

Limitations

Stated up front

Public notional spend, not realized cash. List prices x tokens.
No prompt/completion volume split in the public rankings, so a fixed blend convention with a published band.
OpenRouter public top-50 slice only. No direct-provider enterprise traffic, self-hosted inference, or private terms.
Backcasts use a fixed tariff snapshot and are survivor-selected toward the snapshot's tariff universe; matched-token growth overstates market growth.
Provider-family diagnostics use a versioned slug-prefix taxonomy and remain heuristic.

The recomputation package (raw input snapshots, pipeline code, and revision history) is shared on request via the data page.