Methodology
How we keep the numbers honest.
TokenTally’s estimates are only as good as the data and assumptions underneath them. This page explains how we source provider pricing, normalize it into a comparable format, review changes, and communicate the limits of the resulting numbers.
The goal is not to pretend that AI pricing is perfectly simple. It is to make it understandable enough that a builder or reviewer can follow the logic, verify the sources, and decide whether a scenario is realistic.
Dataset snapshot
43 models tracked
Last synced
3/13/2026
Our dataset lives in a version-controlled workspace so pricing changes can be reviewed before they reach production.
We aim to keep a citation trail for each material pricing revision, including where the number came from and when it was checked.
Providers we monitor
10+ vendors at last sync:
Workflow
Collect
We pull pricing directly from provider pages (OpenAI, Anthropic, Google, Mistral, Meta/Fireworks).
Normalize
All entries are converted to USD per million tokens and tagged with context windows + latency classes.
Review
Each proposed change goes through a human-reviewed diff with source citations before it hits production.
Publish
Our curated dataset drives the calculator UI, preset scenarios, and the comparison table instantly.
Normalization rules
- All costs are normalized to USD per million tokens for both input and output usage.
- Context windows are recorded in raw tokens and surfaced so users can spot unrealistic prompt assumptions.
- Latency classes (economy/standard/premium) are qualitative comparison aids, not guarantees of runtime behavior.
- Calculator outputs are scenario estimates, not invoice-level commitments.
Update cadence
We aim for regular manual review, plus opportunistic updates when providers change pricing, context windows, or product packaging. If a model is deprecated or materially changed, we want that reflected before users rely on stale assumptions.
Quality guardrails
- We try to attach a citation URL and check date to material pricing updates.
- Automated regression tests help ensure calculator changes do not silently alter the math.
- When precision is ambiguous, we prefer to state assumptions clearly instead of implying certainty.
- Editorial pages exist to interpret pricing decisions, not just restate rate cards.
Change management
We are moving toward a clearer change log and evidence trail for pricing revisions so users can understand what changed, where it came from, and when it was reviewed.
Incident response
If a pricing issue slips through, we want it reported quickly and corrected clearly. Send corrections or concerns to support@tokentally.net.