Whisper API Pricing in 2026: What the New gpt-transcribe Changes

If you searched for Whisper API pricing in 2026, the answer changed in late July. OpenAI shipped gpt-transcribe, a new model its docs now call "the recommended model for general file transcription", priced below both whisper-1 and the gpt-4o transcription models. Whisper is no longer the default, and it is no longer the cheapest.

But whisper-1 is still the only OpenAI model that does one specific thing, and the new lineup splits capabilities across models in ways the pricing page does not spell out. Here is the whole picture, verified against OpenAI's own docs in July 2026.

OpenAI transcription pricing, model by model

Three prices matter for most people: $0.27/hour for the new default, $0.36/hour if you need speakers or word timing, and $0.18/hour if you only want the cheapest possible words.

What whisper-1 still does that nothing else can

Word-level timestamps. The timestamp_granularities[] parameter, which returns a start and end time for every word, works exclusively with whisper-1 and its verbose_json response format. None of the newer models support it, including gpt-transcribe.

So in 2026 you face an odd trade: the newest, recommended, most accurate file model cannot tell you when a word was said, and the model that can is the oldest one in the lineup at 33% more per hour.

What gpt-transcribe changes

Launched in late July 2026 alongside gpt-live-transcribe, gpt-transcribe is a 25% price cut against gpt-4o-transcribe with new controls the older models lack: streaming output, multiple language hints, and free-text context for names, jargon, and expected vocabulary. For plain speech-to-text on completed recordings, it is now the obvious pick of the family, which is exactly what OpenAI's docs say.

What it does not change is everything around the words:

• Speaker diarization is not included. Who-said-what requires switching to gpt-4o-transcribe-diarize with response_format=diarized_json, a different model at a higher price with its own parameters.
• The 25MB upload cap still applies to every model. A one-hour meeting recording as WAV blows past it; a multi-hour file in any format does. OpenAI's own guide suggests splitting audio with third-party tools and warns that cutting mid-sentence loses context. Chunking, overlap handling, and timestamp stitching are your code.
• The API returns text and forgets it. There is no transcript storage, no library, no search across past transcriptions. OpenAI's file-search storage product is for documents and vector stores, not a transcript archive.

The real cost: worked examples

Raw API pricing looks unbeatable until you price the whole pipeline. Take a solo podcaster or researcher doing 30 hours of audio a month who wants speakers labeled and searchable archives:

• Raw API route: 30 hours on gpt-4o-transcribe-diarize (you need the speakers, so you cannot use the cheap model) is about $10.80/month, plus you build and maintain chunking for the 25MB cap, storage, search, and a UI. The words cost $10.80; the product around them costs your engineering time indefinitely.
• Product route: a Transcribe.so Starter plan is $9/month for 20 hours a month, with diarization, word-level timestamps, files up to 2 hours (10 hours on Business), a searchable library, and cited Q&A included. Pay-as-you-go is $1 per audio hour with no subscription.

At 10 hours a month the raw API is a few dollars cheaper on paper; at 30+ hours the product is cheaper in absolute dollars before counting a single hour of your time. We walk through the full model-by-model comparison against transcription products in our best transcription tools of 2026 roundup.

When the raw API is the right choice

If you are building your own product or pipeline, gpt-transcribe is genuinely excellent value: $0.27/hour for the recommended model, streaming support, and context hints. Use it when transcription is a component inside something you are engineering anyway, and budget for the chunking and storage work the API leaves to you.

If what you want is transcripts you can use, with speakers, timestamps, and search, a transcription product will beat assembling three OpenAI models plus infrastructure. That is not a knock on the API; it is what "API" means. If you want both, Transcribe.so's API ships the assembled pipeline (diarization, chapters, cited answers, search) as one HTTP call, included on every plan.

Frequently asked questions

How much does the Whisper API cost in 2026?

whisper-1 costs $0.006 per minute, or $0.36 per audio hour. It is no longer OpenAI's cheapest or recommended transcription model, but it remains the only one that returns word-level timestamps via timestamp_granularities[].

Is gpt-transcribe cheaper than Whisper?

Yes. gpt-transcribe costs $0.0045 per minute ($0.27 per audio hour) against whisper-1's $0.006 per minute ($0.36 per audio hour), a 25% saving, and OpenAI's docs recommend it as the default for file transcription. The trade-off: it has no word-level timestamps and no diarization.

Does the OpenAI transcription API include speaker diarization?

Only one model does: gpt-4o-transcribe-diarize at $0.006 per minute, using response_format=diarized_json. It labels speakers A, B, C or matches up to 4 known-speaker reference clips. The new gpt-transcribe model does not support diarization at all.

What are the file limits on the OpenAI transcription API?

Uploads are capped at 25MB per request across all models (mp3, mp4, mpeg, mpga, m4a, wav, webm). Longer recordings must be split client-side; OpenAI's guide recommends third-party tools for chunking and warns against mid-sentence cuts.