Skip to content
AuraMatrix
Use in your AI
← All models

Token Price Index · Team operations

What needs a person right now

The queue below is computed from the data each time the page loads, so it is always current. How to handle each item is in the handbook further down.

Generated 2026-11-06 08:17 UTC119 records, 42 of them Batch / Flex / Priority plans4 unverified14 sources

Promotional price ending or ended 0

Nothing.

Not re-checked for 30+ days 0

Nothing.

Awaiting review 0

Nothing.

Not yet covered 5

ModelChannelRecordDueAction
—meta——Meta has no first-party paid API; Llama is listed through cloud and aggregator channels only.
—Amazon Bedrock (Anthropic models)——Anthropic models on Amazon Bedrock are sold through AWS Marketplace and are not in the Bedrock price list API; not captured yet.
—Amazon Bedrock (regions)——Bedrock price lists for Tokyo, Spain and Frankfurt do not list Llama 4 Maverick; only us-east-1 is shown.
—mistral——Mistral's Batch and Priority price tabs render client-side; only the Standard tab is captured so far.
—Vertex AI (service plans)——The Cloud Billing Catalog lists Flex, Priority and Off-Peak SKUs for Gemini; only Standard and Batch are recorded so far.

Handbook

This page and the handbook are maintained with the code. When a procedure changes, change it here; do not keep a separate document.

Cadence (trial, adjusted as we learn)

  • Every Monday: the weekly sweep (below). Unchanged sources only update the sweep time; changed prices get a new record.
  • Daily: the cloud price APIs (Azure, AWS, GCP) are part of the same sweep script and can be run alone on weekdays.
  • Within 48 hours of a vendor announcement: add the new record. A promotional price shows up under “ending” 14 days before its end date.
  • Review within 7 days: an independent second reader re-reads each new record within a week (blind re-read by Codex, below; people spot-check). Records not re-checked for 30 days are marked stale on the public pages, never hidden.

The weekly sweep, step by step

  1. On the server as admin, run dev/claude/price-tools/sweep.sh. It fetches all 13 sources into dev/claude/price-snapshots/<date>/, records SHA-256, size and fetch time, and prints one line per source: unchanged, CHANGED (with the new dollar amounts it saw), NEW or PARSE ERROR.
  2. For every CHANGED source: open the page or API response in that folder, find which model and channel moved, and add a new record to quotes.json with observed_at equal to the sweep's FETCHED_AT and a new entry in sources.json pointing at the new snapshot. Never edit the old record.
  3. For a PARSE ERROR: the page layout changed. Note it under coverage gaps rather than guessing, and tell Claude so the parser is fixed.
  4. Refresh the usage panel: python3 -I dev/claude/price-tools/import-openrouter-usage.py dev/claude/price-snapshots/<date> (it reads or-rankings.html from that folder).
  5. Update last_sweep_at in meta.json to the sweep's fetch time, run node --import tsx --test tests/prices.test.ts, commit on claude/public-site. Claude builds and deploys.
  6. Snapshots stay on the server and must be archived to R2 before any are deleted.

Reviewing records: blind re-read by Codex, spot checks by people

  1. Generate the tasks: python3 dev/claude/price-tools/verify-tasks.py <label> writes dev/claude/verify/<label>/tasks.json and PROMPT.md. Tasks name the snapshot file, model, channel and plan but carry no recorded prices. Aggregator and reseller channels are included. Records whose price was derived from a sentence on the page (not listed as a number) are routed to a person instead.
  2. Run the reader: codex exec --sandbox workspace-write -C dev/claude/verify/<label> "$(cat dev/claude/verify/<label>/PROMPT.md)". It reads each snapshot on its own and writes results.json. It must not open data/prices/ or the live page.
  3. Compare: python3 dev/claude/price-tools/verify-compare.py <label> (dry run), then with --apply. Exact matches become VERIFIED (operator, cloud and reseller channels) or LISTED (aggregator channels) with reviewed_by = codex (agent) and reviewed_at; everything else lands in MISMATCHES.md.
  4. A person handles MISMATCHES.md and the human-routed tasks: open the source link, compare, and either confirm (set reviewed_by to your name) or tell Claude the number you see. A changed page is a price change, not an error: the old record stays and a new record is added with a new snapshot.
  5. Spot checks: each week a person re-checks a handful of VERIFIED records against the live page and notes the result here. The reviewer can never be the observer who recorded the record. Marketplace (aggregator) listings that match become LISTED, never VERIFIED; the checks refuse VERIFIED on them.
  6. Commit. Claude builds and deploys, and the queue here shrinks.

Adding a record

  1. Fetch the source page or API, save the raw response to the snapshot directory, note its SHA-256, size and fetch time.
  2. Add the source to sources.json (or reuse the source from the same fetch) and the record to quotes.json. observed_at must equal the source's fetch time.
  3. Units are per 1M tokens in the channel's own currency. Fill cached and tier prices when the source lists them; leave them out otherwise, never write 0. A Batch, Flex, Priority or Ultrafast price is its own record with service_plan set, never a field on the standard record.
  4. Run the tests. A record without a source, a reviewer equal to the recorder, or a VERIFIED aggregator price is refused.

Rules

  • Records are appended, never edited or deleted. A correction is a new record.
  • The reference price comes only from official channels: a verified one if there is one, otherwise the newest, shown as unverified.
  • No number is shown without a source. Missing values are a dash, not 0 and not free.
  • No resellers, no contract prices, no tax. Aggregator prices are labelled and never used as the reference.
  • Snapshots live on the server under dev/claude/price-snapshots/. They are not yet archived to R2; do not delete them before that.

Contact

Corrections: hello@auramatrix.ai. Code and data are on the claude/public-site branch; Claude deploys with am-deploy-public-site.