Signal / AI · speech-to-text / Checking AI transcription costs

Why your AI transcription cost report can rise without a price rise

LiteLLM v1.108.15 starts recording listed costs for Groq and Meta transcription that its earlier log showed as $0. If you check transcription costs, keep the gateway estimate separate from what the provider bills.

Use this if

You already use LiteLLM to track Groq or Meta transcription. The worked example applies only to Groq Whisper Large v3 Turbo at its standard published rate.

No action if

If you cannot match the provider, model and billing mode, do not infer a bill increase or change your budget from the gateway alone.

Article time zone:

Asia/Singapore

Who is outside the scope:

No customer recordings, credentials, production changes, or paid test required. No logged test result, actual invoice, provider price rise or historical backfill verified.

This compares reporting and charges only. The release also lists a separate change for setups without a provider key.

First published

AI

OTR Emerging Data

AI transcription · cost tracking

The cost report can rise without a price rise

An AI gateway can report a higher cost even if the transcription price stayed the same.

What the sources show

PyPI recorded the LiteLLM v1.108.15 source archive on 22 November at 05:34:40 SGT. The separate GitHub long-term support release says it fixed $0 spend records for Groq and Meta transcription. Groq lists US$0.04 for one audio hour on Whisper Large v3 Turbo.

The new connection

One ten-second clip and two five-second clips contain the same amount of audio. With Groq’s ten-second minimum per standard request, the latter represent twenty seconds of estimated billing, not ten.

Same audio length, different billed seconds
  • One ten-second request10 seconds
  • Two five-second requests20 seconds
Groq published minimum; illustration only.
  1. The old LiteLLM log showed zero.
  2. The new version starts recording listed cost.
  3. Compare the log with approved usage and the provider bill.

A realistic next move

Pull one known transcription from your approved test records. Compare the gateway amount with the published price for the exact call. Use the next provider bill to confirm what was charged.

The illustration is not a measured LiteLLM result or a customer bill. Batch discounts, credits and contract terms can change the invoice.

The connection

A recorded zero is not proof of free transcription

The new LiteLLM LTS release records listed costs for Groq and Meta transcription where the older log showed zero. PyPI shows the v1.108.15 source archive was uploaded at 05:34:40 Singapore time on 22 November.

Reporting change, then billing check
  1. Before releasePrior logs could show zero
  2. 22 Nov 05:34:40 SGTv1.108.15 source archive upload
  3. Next statementReconcile matching charge
No provider price rise inferred.
Observed

The release and upstream change are publicly documented. We have not operated a customer gateway or inspected real billing.

The connection

If your old budget comes from the zero in the software report, check the provider’s bill before using it as a baseline.

How the effect travels

The gateway estimates cost. The separate provider bills for usage under its own minimums and rates.

The upstream change documents a prior zero entry, not a prior zero provider charge.

The decision path

Check one audio route before acting

1. Find one known test

Use approved records and the exact model and route.

2. Count how it was sent

Record audio seconds, request count and batch versus standard.

3. Compare estimates and actual charge

Use the correct public rate for an estimate, then use the provider statement for actual billing.

The example uses public dollar amounts from Groq and does not extend to all plans or to Meta.

Optional research / Checks 1–10

What the sources support

Money flow

Project estimate is not provider cash.

Orders

No paid order or deployment verified.

Bottleneck

Match the same model, provider and request mode.

Beneficiaries

Teams already tracking Groq or Meta.

Company finances

No customer spend or provider revenue found.

Costs payback

Use seconds and per-request minimum before pricing.

Ownership

LiteLLM records estimate; Groq bills separately.

Hiring

No hiring claim.

Government

No government approval.

Language

English first-party sources.

Optional research / Checks 11–20

What would change the reading

Behaviour

Review an existing approved record.

Contradiction

Estimate can rise while invoice does not.

Missing proof

No settled customer invoice or repeated live test.

Approval

No approval, account or purchase required.

Alternatives

Compare batch terms and credits.

Geography

No assumption of worldwide account terms.

History

Earlier gateway log showed zero, not a zero bill.

Awareness

Available package is not evidence of adoption.

Countercase

Different rate or batch may explain a mismatch.

Action clock

Reconcile on next actual provider bill.

Evidence & next review

Read the originals

PyPI source archive upload

Published 2026-11-21T21:34:40Z

PyPI records this source archive upload on 21 November at 21:34:40 UTC, or 22 November at 05:34:40 SGT. This is a package upload, not the GitHub announcement clock or the first wheel upload.

Source record

Publisher: Python Package Index

Original language: English

Source publication date 2026-11-21T21:34:40Z

The official GitHub release page separately displays 21 November 21:35. Its note confirms the spend-recording change; the PyPI source archive provides the exact package clock.

Source record

Publisher: LiteLLM

Original language: English

Documents prior zero spend and the provider cost map change.

Source record

Publisher: LiteLLM

Original language: English

Describes Groq and Meta transcription spend now recorded at listed rates.

Source record

Publisher: LiteLLM

Original language: English

Groq lists $0.04 per hour; this is not customer invoice evidence.

Source record

Publisher: Groq

Original language: English

Lists a 10-second minimum billed length per request, conditions and scope.

Source record

Publisher: Groq

Original language: English

Lists half the synchronous API rate; do not treat standard pricing as universal.

Source record

Publisher: Groq

Original language: English

No invoice, production test or deployment evidence. PyPI proves archive availability, not adoption. Groq’s public standard rate is not an account-specific price or a Meta rate.

Consequences and economics

First-order consequence

The released gateway changes how it records two providers’ transcription costs.

Second-order consequence

A cost report can rise as previously missing estimates appear.

Third-order consequence

A team can reconcile new records against an unchanged like-for-like usage sample.

Fourth-order consequence

If the mismatch remains, separate minimums for each request, batching and account-specific terms before escalating it.

Fifth-order consequence

Only change forecasts if repeated matching usage and actual bill evidence justify it.

Observed costs and benefits

At $0.04 per hour, the listed standard estimate is about $0.000111 for one ten-second request and $0.000222 for two five-second requests billed for ten seconds each, before credits or rounding.

Challenge and uncertainty

Disconfirming evidence

Batch prices or customer credits can differ; avoid treating the illustration as an invoice.

What would confirm it

A provider usage record or invoice matched to the specified route confirms actual billing; the project release alone does not.

What would disprove it

If identical paid usage shows no provider bill change, a new non-zero gateway log entry does not establish a price rise.

What proves what?
What proves what?
RecordWhat it can show
Gateway logEstimated spend under its rules
Groq model pagePublished standard hourly price
Provider statementActual billed amount for matching usage
Available records are not results from this report.

AI powered research