THE METHODOLOGY DEPARTMENT

HOW THIS SITE KNOWS

One brand, two tones. The editorial pages are dry and ironic. This page is neutral and precise — it is the record of how every number on the site is produced, labeled and verified.

Classes of evidence

Every surface on this site falls into one of four classes, and the name tells you which one. The word Index is reserved for artifacts that are calculated or based on evidence — never for editorial commentary.

What “verified” means

Every model on the board carries a verified date, and it means one specific thing: the last day an automated run confirmed that price against the daily source — the LiteLLM community dataset, with a pinned key per model and sanity guards. It is not a human sitting down with the vendor's pricing page. The sourceUrl next to it points at the provider's own pricing page, which is the authority for the price — but it is not what the daily job reads.

The date is written by the watcher, per model, on every run that confirms it. A model whose check fails — its key vanishes from the source, or the change trips the 3x sanity guard — keeps its old date. So a stale date is not an oversight: it is the site telling you we have not been able to confirm that number lately, and after 3 days the model page says so in as many words. Machine-readable consumers get the same field in pricing.json.

Known limits, stated rather than buried: the cross-check inherits whatever lag the community dataset has against the vendors' own pages, so a price can be wrong here and “verified” on the same day if the source itself is behind. Until 2026-08-12 this field was stamped by hand and did not move on its own; it now updates daily, and its history is auditable in the repo's commit log.

Cache, batch and reasoning rates

Headline input and output prices are only the part everyone quotes. What a repeated workload actually costs depends on the rates underneath: cached input, cache writes, batch tiers, separately-billed reasoning tokens, and the minimum prompt length below which caching never kicks in. Those live in a machine-owned file the daily watcher regenerates from the source in full — no hand editing, and no site-wide convention standing in for a provider's published number.

When a provider publishes no rate for something, we print “not published” instead of a number. That is not pedantry: until 2026-08-12 this site derived cached input as a flat 10% of list and cache writes as 1.25x list, which is exactly right for Anthropic, OpenAI and Google — and badly wrong for DeepSeek, whose cached reads cost about 1% of list. We were publishing a cached DeepSeek price twelve times higher than reality, and inventing a cache-write fee for every model whose provider charges none — 15 of the 36 on the board today. Both errors fed straight into canonical workloads. Corrected, the cached-chat profile moved by up to 13%.

Consequences of the rule, stated plainly: where a provider publishes no cached-input rate, a workload is computed without cache rather than credited with a discount we cannot verify — so its cost goes up, not down. Where no cache-write rate is published, writes are billed at zero. Both cases are labeled on the page and flagged in pricing.json, where null always means “not published” and never “free”.

That rule had a blind spot until 2026-09-23, and it is worth stating because it is the mirror image of the error above. “Our source has no number” is not the same claim as “the provider publishes no number”, and the page was making the second claim while only entitled to the first. Claude Opus 5.5 joined the board with a batch tier printed in Anthropic's own pricing table that the daily source did not carry at the time — so its page announced “no published batch tier for this model” about a rate the provider publishes in public. The source has since caught up with Anthropic's batch rates and the declarations were removed, which is the loop working: a declared discrepancy is meant to expire, not to become furniture. Two things changed: labels now say what our source holds rather than what a provider supposedly offers, and a rate we can see on the provider's page but not in the source can be declared as an override, with its URL and the date we declared it, shown on the model's page. A declared discrepancy beats a confident absence.

How estimates are made

Costs at list prices are computed as (input × inputPerMTok + output × outputPerMTok) / 1M against the verified public prices on the pricing board. Where a usage file reports cached reads, they are billed at ~10% of the input list price (the prompt-caching convention documented in our prompt-caching guide). Everything is an estimate at list prices — never your invoice, never your negotiated rate.

Canonical workloads apply the same math to four versioned, fully-declared profiles (chat, cached chat, coding agent, batch). Every assumption is published with the profile: calls per month, tokens in/out, cache hit and write share. A workload is a cost profile, not a performance benchmark — and there is no universal "real cost".

The burn score and compression in the calculator are heuristics over text (filler, repeated 4-grams, ornament, prompt size vs a baseline). They judge verbosity, not correctness. The compressed prompt is labeled “heuristic — review before use”.

Evidence levels for waste

“Token burn” is spend that did not contribute to the accepted result. We never publish a single universal efficiency score, and we never compare burn between organizations that use different baselines. Instead:

Sources and versioning

Prices are checked daily against public list prices by an automated watchdog (pricing-watch), with sanity guards (a >3× change opens a review instead of applying). Each change is recorded with its date, direction, and before/after in the Price Changes history. The versioning exists so that, months from now, this site can answer: “what did Claude, GPT and Gemini actually cost for workload W in August 2026?” — the historical series is an asset, not a byproduct. Board last changed: 2026-10-13.

Data policy

What we collect: aggregated pageviews and events via PostHog (EU-hosted), used only to understand what people actually use. Newsletter signups send an email to our own list and nothing else.

What we do not collect: the CSV you audit in Audit My Month never leaves your browser — the whole analysis runs locally. Tools never upload prompts, exports or identifiers. No personal data is sold or shared.

Opt-out: append ?tbx-optout=1 to any page URL to disable analytics in your browser (revert with ?tbx-optin=1). Aggregated stats are never reported with sample sizes too small to be meaningful (see Privacy).

Conflicts of interest

TokenBurn Index sells no tokens, accepts no affiliate commissions and takes no money from model providers. Pricing comes from public list prices, not vendor relationships. If a paid audit service ever exists, every piece of data it produces will carry the evidence class and the cohort behind it. When an article has any relationship with a vendor, it is disclosed in the article.

Every feature is judged against one question: does it strengthen the historical data asset, methodological credibility, current useful information, organic discovery, shareability, or repeat usage? A tool that only exists to be a tool is deferred.

Limitations

Estimations are not invoices. Token counts use the o200k tokenizer — exact for GPT models, an approximation elsewhere. We used to quote ~±10% for other vendors and no longer do: Anthropic states its 4.7+ models turn the same text into roughly 30% more tokens than its earlier ones, which is wider than that band claimed. An error bar we cannot substantiate is worse than admitting we do not have one — see the tokenizer tax. Negotiated rates, discounts and free tiers differ from list prices. Energy disclosure is only ever scored as “is clear information available?”, never as an estimated carbon figure — we do not have a serious methodology for that, so we do not fake one.

Version: methodology v0.1 · Applies to surfaces published from 2026-08-12 onward.