Whisper V3

Whisper V3 is OpenAI's open-weight speech family around large-v3, not the current hosted Speech-to-Text default. Rechecked 2026-08-17: github.com/openai/whisper had 107,437 stars, license MIT, last push 2026-07-28. Treat this page as the open repo, not whisper-1 on the API and not whatever newer hosted transcription SKU OpenAI is selling now. Drop leftover latest STT / English WER <3% / most powerful open-source as an audited 2026 crown.

Compare whisper-1 if you wanted the older API object, OpenAI if you wanted the live hosted catalog, and Deepgram Nova-2 if that page is still the leftover hosted alternative.

Key Features

• Open weights, not the API default: The GitHub repo is MIT inference plus model cards. Recheck OpenAI's current transcription model list before you send production audio to "whisper-v3" as if it were the hosted SKU.
• Family, not one file: Leftover copy listed Tiny through Large-v3 sizes. Recheck the repo README for the names that still download.
• No invented WER: <3% English WER is out unless you re-run a first-party card. We did not.
• Not a 99-language census we re-counted: The original paper and cards talk multilingual training. Recheck the card; do not freeze 99 as law.

Use Cases

• People self-hosting Whisper weights who need the MIT repo, not a hosted default.
• People comparing leftover large-v3 to the current OpenAI transcription API.
• People who were about to paste WER <3% into a 2026 deck.

Limitation: GitHub raw README was rate-limited in this pass. Stars/license came from the API. Recheck the README locally.

Pricing

GitHub on 2026-08-17. Hosted audio is a different bill.

If a leftover blog still says Whisper V3 is OpenAI's latest hosted STT, treat this table as the source.

Getting Started

1. Open github.com/openai/whisper.
2. Recheck the current OpenAI transcription model list before you use the API.
3. Recheck whisper-1 if you meant the older API object.
4. Do not paste WER <3% into a deck.

Frequently Asked Questions

Is this the current OpenAI speech API?

Not something this page claims. This is the open-weight repo.

Is it still the most accurate open STT?

We did not re-run a bench. Do not keep that crown.

Same as whisper-1?

whisper-1 is the leftover API object page. This page is the GitHub family.

Alternatives

• whisper-1: Older OpenAI API object.
• OpenAI: Lab and live API catalog.
• GPT Image 2: Unrelated stills model if you followed the wrong OpenAI page.

Tips

1. Quote MIT and 107,437 stars from the GitHub check.
2. Split open weights from hosted STT.
3. Recheck the live API model name before you write a contract.

Conclusion

Whisper V3 is the MIT open-weight large-v3 family with 107,437 stars, not OpenAI's verified 2026 hosted STT default. Start at github.com/openai/whisper, then decide whether the current OpenAI transcription API already covers the job.