Inference APIs
Blog/speech, benchmarks

OpenAI's Whisper API against ours, same files, timed

Three audio files, three runs each, the same afternoon: OpenAI's whisper-1 took 13.3 seconds on a four-minute recording, Whisper Large v3 here took 2.9, and the transcripts matched at 99.6 percent word for word. Price is $0.006 against $0.002 a minute. The numbers, the method and the caveats.

Taylor Hawkes · November 18, 2026

Most of the people who emailed us about transcription in September were running OpenAI's whisper-1 in a dictation app, a meeting-notes tool or a subtitle pipeline, and the question was always the same: is the open Whisper on a different host actually faster and cheaper, or just cheaper? So on 22 September 2026 we timed both on the same files from the same machine, three runs each, back to back.

Wall time, upload to transcript

FileOpenAI whisper-1, 3 runsmedianInference APIs whisper-large-v3, 3 runsmedian
8-second dictation clip (WAV)1.8 s, 2.8 s, 1.7 s1.8 s1.1 s, 1.5 s, 0.8 s1.1 s
33-second speech (MP3)14.6 s, 12.1 s, 12.8 s12.8 s17.8 s, 3.3 s, 2.5 s3.3 s
4-minute audiobook chapter (MP3)20.0 s, 13.3 s, 10.5 s13.3 s2.7 s, 2.9 s, 5.8 s2.9 s

The 8-second clip is dominated by upload and model start, so the gap is small. On anything longer OpenAI's time grows with the audio and ours mostly does not: the four-minute file came back in under three seconds on two of three runs. One run on our side took 17.8 seconds (the 33-second file, first attempt), which we include rather than hide; it was the cold start of a GPU that then served the next two runs in about 3 seconds.

Same transcript

On the four-minute chapter both transcripts were 927 words long. A word-level diff found three differences: "size" against "sized", "hand searches" against "hands searching", and "to" against "too". That is 99.6 percent identical, and reading the audio, ours was right on two of the three. Punctuation and sentence breaks differed slightly ("Chapter 35, London Heathrow." against "Chapter 35 London Heathrow"), which is normal between Whisper versions.

Price and features

OpenAI whisper-1Inference APIs whisper-large-v3
Price per audio minute$0.006$0.002
An hour of audio$0.36$0.12
100 hours$36$12
Word timestampsverbose_jsonverbose_json, same shape
Speaker labelsnot offereddiarize=true, included
Modelwhisper-1 (OpenAI's hosted Whisper, large-v2 lineage)Whisper Large v3, open weights

How to switch

The whole change for an OpenAI SDK user
client = OpenAI(
    base_url="https://api.inferenceapis.com/v1",   # add this line
    api_key=os.environ["INFERENCE_API_KEY"],
)
with open("meeting.mp3", "rb") as f:
    text = client.audio.transcriptions.create(
        model="whisper-1",          # accepted as an alias; or "openai/whisper-large-v3"
        file=f,
        response_format="verbose_json",
    )

Everything else in the call stays as it is: the multipart upload, language, prompt, response_format including srt and vtt, and timestamp_granularities. The translations endpoint is not offered; transcribe in the spoken language and translate the text with a chat model on the same key. Audio is processed in the United States and not retained.

Caveats

  • One afternoon, one location. Both services vary with load. We measured from a residential connection in the United States; your numbers will differ, and the right test is your own files from your own servers.
  • Different Whisper versions. OpenAI's whisper-1 is its hosted Whisper from the large-v2 lineage; ours is Large v3 with open weights. On English read speech they agree almost entirely; on other languages the per-language accuracy pages show where Large v3 is strong and where it is not.
  • Cold starts happen. The 17.8-second outlier above is what one looks like. If a first request after idle time is slow, the next one is not.

Prices and the endpoint reference are on the speech-to-text API page; the side-by-side with other hosts is at Whisper Large v3 vs OpenAI whisper-1.

More posts