Methodology & data.

How a row is built

One row is a product × plan × capture date. A product qualifies when its public pricing page carries a numeric, depletable meter — credits, tokens, tasks, actions, outcomes. Because a credit means something different everywhere, the Index works in three comparable layers and never compares raw $/credit across products:

• Unit-free within-product ratios — top-up premium over a plan’s own bundle, tier-to-tier discount curves. The unit cancels.
• Policy distributions — rollover, expiry, behaviour at zero, seat scaling. Counted across the panel.
• Task-equivalent prices within a category — only where the credit-to-action mapping is public, and only where the task is the same.

Sources are vendors’ own public pricing pages and help docs only — never third-party trackers, leaks, or estimates (those may annotate a row, never populate it). Every published row is pinned to at least one archived copy of the vendor’s page: a Wayback Machine snapshot, or — where the vendor blocks Wayback or the page renders only in a browser — a dated screenshot hosted with the Index. Promos and regional prices are flagged and excluded from metrics. Prices visible only behind the logged-in purchase screen are recorded as in-app only and kept out of public aggregates — the opacity is reported, not imputed. Derived figures are computed at build; none is hand-entered, and missing is missing — no imputation, ever. One dated editorial exclusion applies to the rate aggregates only: Bolt meters individual tokens, so its bundled rate (as low as $0.0000018 per token) sits about 90 times below the next-lowest rate on the panel and nearly four orders of magnitude below the median, and it alone drove the Edition 1 whole-panel spread to 13 million to one. Excluded from rate aggregates from Edition 2 (decision 2026-10-12). The product stays on the panel for every other count and in every export row.

What we excluded — and why

The inclusion test is deliberately strict, and plenty of well-known AI software fails it. The reasons are part of the picture — not every AI product meters, and where the meter is absent is itself a finding:

• Flat per-seat AI, no meter — e.g. Granola (AI meeting notes; $14–$35/user, “unlimited”). Priced like classic SaaS — nothing to meter.
• Caps without purchasable units — e.g. Wispr Flow (AI dictation; 2,000 words/week free, then a flat “unlimited” Pro at $15). A cap, but you upgrade rather than buy more, so there’s no top-up economics to read.
• No first-party AI product — e.g. Obsidian (a notes app; AI is third-party plugins). Out of scope.
• Raw API token pricing (OpenAI / Anthropic / Google per-token rates) — a different layer; this study is about the abstraction products wrap around tokens.
• Negotiated enterprise pricing — unobservable; tracked instead on the opacity watchlist.
• Promos, regional, and lifetime prices — flagged and kept out of the metrics so the quarterly time series stays clean.

The panel & source ledger

Every published product, its computed bundled-rate range, whether the top-up price is public, and the dated source + archive each row rests on. The full per-plan dataset is the /benchmarks/credit-index.json endpoint (published observations only). Relay.app shut down in September 2026; it is out of the current view and its calculations, and stays in the editions observed before the exit (its page).

The Index is published under CC BY 4.0: free to reuse — including by AI engines — with attribution to Strategy of Finance — The AI Credit Index.

† Borderline inclusion — sits near the raw-API exclusion line; captured anyway, flagged here.

Field definitions

The JSON and CSV exports share one row shape — one plan tier per row, with the product and observation fields repeated on each row. Computed fields are marked; they are derived at build time from the published inputs and are never stored. The same definitions are kept in the repository as docs/ai-credit-index-data-dictionary.md.

Downloads: current view JSON · CSV · Edition 1JSON · CSV · Edition 2JSON · CSV. Every export carries its view id, observation window, generation time and licence in its metadata (JSON) or response headers (CSV).

Corrections and updates

A published figure is corrected by adding a dated observation or a dated correction note — never by silently editing a record. Corrections inside an edition window are listed on that edition’s page (Edition 1, Edition 2); coverage and methodology changes are logged on the change log. To report an error, use the contact form with the product, the figure and the vendor page you read.