Start with the primary source.
Provider mechanics and pricing claims link to the provider’s own documentation and remain scoped to the relevant model or API.
TokenGauge keeps provider pricing, application attribution, evidence strength, and experiment limits visible. Reconcile a bill locally, map spend to customers and features, then test one controllable cost bucket.
01Paste aggregatesInput, cached-input, output, attempts, accepted answers, and the provider bill.
02Reconcile the rateSeparate modeled token spend from a diagnostic invoice gap.
03Test one changeTake the largest controllable bucket into a quality-gated experiment.
Your entered usage and bill values stay in the browser. Reconcile the bill, attribute it with the free ledger, or request a fixed £75 implementation for one authorized workflow.
The provider knows which model and token buckets were billed. Your application knows which customer and feature caused the call. The guide shows how to join those ledgers with opaque labels, count retries, price the correct rate window, and leave invoice residuals visible.
Use stable opaque customer, project, and workflow labels—never email addresses or prompt content.
Keep uncached input, cache reads, output, attempts, and accepted answers attached to the effective rate card.
Compare request and token counts first, then explain credits, tools, storage, commitments, taxes, and other residuals.
The dated live directory contains 58 pricing cards across nine provider families. A card may represent a context band, region, cache mode, or effective-date variant—not a distinct model.
Rates are dated and source-linked. Provider-specific context thresholds, regions, cache modes, effective dates, tool fees, and storage charges are not flattened into a universal comparison.
The maintained catalogue contains 120 evidence cards, including 12 open cards. A separate 2,500-row atlas contains 1,316 atomic candidates and 1,184 compound configurations. Neither count is presented as a count of distinct proven methods.
Provider mechanics and pricing claims link to the provider’s own documentation and remain scoped to the relevant model or API.
Break-even math and comparisons are labelled as analysis, not repackaged as provider promises.
Supported lab recipes are labelled individually. Other cards remain guided protocols until a real adapter exists.
A smaller token count only wins if quality, retries, fallbacks, tools, latency, and rejected answers still produce a lower cost per accepted result.
The public rate directory and a 12-row atlas sample remain useful without a purchase. Paid tiers add the complete evidence catalogue, server-filtered research atlas, encrypted provider connections, and measured experiment tools without bundling API credits.
The first 100 authenticated identities receive £5 Pro, £15 Pro+, or £20 Ultimate pricing. Check the maintained app for the authoritative live place count.
Continue with ChatGPT or emailBrowse the current snapshot and open catalogue in the live app. Supported A/B recipes are identified individually and remain available after connection.
Go to the live applicationThe complete evidence catalogue, all 2,500 research-atlas rows, all nine encrypted provider connections, and a 100-row experiment dashboard.
Review ProEverything in Pro, a 250-row dashboard history, and experiment CSV export.
Review Pro+Everything in Pro+ with a 1,000-row dashboard and CSV export window for repeat testing.
Review UltimateA separate TokenGauge password is optional. Continuing with ChatGPT does not itself charge you; Stripe checkout remains a separate step, and bring-your-own-key requests are billed by the selected provider.