Modeled token spend
Uses the selected input, cache-read, and output rates. It excludes tools and any billing category you did not enter.
FREE AI API BILL AUDIT
Find the cost bucket before changing the model.
Bring aggregate token totals and your provider bill. TokenGauge reconciles 52 dated rate cards, cache share, accepted-answer economics, and an explicitly approximate retry burden—entirely in your browser.
Rate snapshot verified 2026-08-16 · no login · no usage upload
BILL RECONCILIATIONSeparate tokens,
Separate tokens,
quality, and variance.
Select the exact provider rate card, copy aggregate usage totals, and compare the token model with the reported bill. A remaining gap is a question to investigate—not an automatic overcharge claim.
Browser-local calculation. Token totals and bill values are not submitted to TokenGauge or stored.
AGGREGATE BILL RECONCILIATION
OpenAI · GPT-5.6 TerraStandard · up to 272K input: $2 input · $0.2 cache read · $12 output / 1MGlobal · choose the rate card that actually covered each request
Modeled token spend$103.80
Reported provider bill$110.00
Bill above token model: $6.20 (5.6%)
MODELED COST DRIVERS
- Uncached input
- $42.00
- Cached input
- $1.80
- Output
- $60.00
- Cache share
- 30.0%
Output is the largest modeled token-cost bucket. A published cache-read rate is applied only to the cached-input bucket.
QUALITY AND RETRIES
Accepted-answer rate85.0%8,500 accepted of 10,000 attempts
Modeled cost / accepted answer$0.01$0.01 per attempt
Uniform-attempt estimate: $15.57 of modeled spend sits on non-accepted attempts. Real retry cost needs retry-specific token buckets.
The variance is a diagnostic gap, not proof of overbilling. Tools, cache writes/storage, images, audio, regional or priority uplifts, taxes, credits, rounding, and mixed price bands can explain it.
OpenAI API pricing ↗READ THE RESULT CAREFULLY
Three numbers. Three different decisions.
Invoice variance
Flags the unexplained gap for investigation. Mixed tiers, media, cache writes, credits, or tax may account for it.
Accepted-answer cost
Stops cheap failed attempts from looking efficient. Validate changes against a declared quality bar.