Speech-to-Text APIs: 14 Models, $0.002 to $0.016 a Minute

A speech-to-text API costs between $0.002 and $0.016 a minute at list rates, an 8x spread for what sounds like the same job, and the cheapest tier is not where most comparisons point. We run 14 transcription models behind one gateway, from OpenAI’s token-billed gpt-4o family to a China ASR tier (ByteDance, Alibaba) that international roundups routinely skip. This page is the category map: what the billing shapes mean, what every model lists per minute, and how to pick by scenario rather than by brand.

TL;DR

• Per-minute list rates across the 14 models span $0.002 (seed-asr, paraformer, fun-asr-realtime) to $0.016 (Google’s Chirp), an 8x spread.
• OpenAI’s gpt-4o transcription models bill tokens, not minutes; on our measured audio they converted to about $0.0031/min (mini) and $0.0062/min (full).
• The cheapest tier is Chinese ASR: ByteDance seed-asr and Alibaba’s paraformer/fun-asr/qwen3 families all list at $0.002-$0.0021/min with streaming variants.
• On clean speech, accuracy is not the differentiator; our seven-model test measured ~0% error on English, Spanish, and French for every model that accepts the language.

What does a speech-to-text API cost in 2026?

The honest answer is a table, because billing shape matters as much as the rate. Per-minute models charge duration regardless of content; token-billed models resample your audio and charge input plus output tokens, which converts to an effective per-minute figure that moves a little with speech density:

List rates are live on each model page; the two ≈$/min figures for the gpt-4o models are what our own multilingual test audio actually billed through the gateway, from the seven-model transcription study. The batch models above are callable through the same OpenAI-compatible /v1/audio/transcriptions endpoint with one API key (we verified this includes the diarize model); the -realtime variants are streaming models with their own interface, documented on each model page.

How does token billing convert to per-minute cost?

Token-billed transcription scales with spoken content, not file size, and lands mid-pack in practice. The gpt-4o transcription models resample audio before tokenizing, so a heavy 320 kbps MP3 and a lean 16 kHz WAV of the same words cost about the same; compressing files saves storage, not transcription spend. On normal-paced speech our measurements came out near $0.0031/min for gpt-4o-mini-transcribe and $0.0062/min for gpt-4o-transcribe; denser or quieter audio moves those a little in either direction. If your budget needs a hard ceiling per audio-hour, a per-minute model gives you one; a token-billed model gives you a forecast.

What is the China ASR tier, and why is it missing from most comparisons?

The cheapest and most streaming-rich block of the table is Chinese ASR, and most English-language roundups never test it. ByteDance’s seed-asr-bigmodel is the outright floor at $0.002/min. Alibaba ships three families: paraformer (streaming-first), fun-asr (batch, flash, multilingual mtl, and realtime variants), and qwen3-asr (flash and flash-realtime), all at $0.002-$0.0021/min. That is a third of whisper-1’s rate and an eighth of Chirp’s.

Two caveats from our measurements keep the picture honest. First, seed-asr has to be told the audio’s language up front; it wins when you know the language, while qwen3-asr-flash at $0.0021 is the cheapest model that auto-detected every language we tested. Second, cheap does not mean slow: qwen3-asr-flash returned in about 3 seconds on our clips, comparable to the OpenAI models. For Mandarin-heavy workloads this tier is not just cheaper; it is the home-turf option.

Which model should you pick for your workload?

Pick by billing shape and streaming need first; accuracy on clean speech will not separate these models. Scenario shortlist:

If the job is a two-way voice conversation rather than transcription, that is a different product with different economics; see our GPT Realtime pricing study for speech-to-speech costs.

FAQ

How much does the Whisper API cost?

whisper-1 lists at $0.006 per minute of audio, billed on duration alone; a one-hour file is $0.36 regardless of format or bitrate. Repetition earns no discount (we re-sent the same clip five times and paid the same price each time), and half the table now undercuts it: the Chinese ASR tier at $0.002-$0.0021/min runs a third of whisper’s rate, and gpt-4o-mini-transcribe measured about half.

Is a more expensive speech-to-text model more accurate?

Not on clean speech. In our seven-model test, everything that accepts the language scored roughly 0% error on English, Spanish, and French; the differences that matter live in cost, streaming, language coverage, and features like diarization. Hard audio (accents, noise, code-switching) is where you should run your own sample, and the transcription study shows where cracks first appear (Mandarin and Hindi in our test).

Which speech-to-text APIs support streaming?

Two shapes exist. The gpt-4o transcription models stream token-by-token text over the OpenAI API shape. The Alibaba realtime family (paraformer-realtime-v2, qwen3-asr-flash-realtime, fun-asr-realtime) streams at per-minute rates from $0.002/min. whisper-1, seed-asr, and the Chirp models return complete transcripts only.

All 14 models run behind the same gateway endpoint with unified billing; current list rates for every model are on the live models page, and the measured seven-model comparison, latency and accuracy included, is in the transcription cost study.

← Back to blog