Models

Speech

Transcription and synthesis, on OpenAI's own endpoint paths. Priced in their own units rather than converted into tokens, because that is how they are metered: you are billed for the seconds of audio you send, or the characters of text you send.

Model	Maker	Does	Price	Endpoint	Licence
whisper-large-v3-turbo	OA OpenAI	speech to text	$5 / M seconds	/v1/audio/transcriptions	MIT
qwen3-asr-1.7b	Q Qwen (Alibaba)	speech to text	$11.25 / M seconds	/v1/audio/transcriptions	Apache 2.0
kokoro-82m	HG Hexgrad	text to speech	$0.93 / M characters	/v1/audio/speech	Apache 2.0

Transcription is rounded up to the whole second, so a three-minute voice note costs about $0.0009. Repeated synthesis of the same text in the same voice is served from cache at a quarter of the price.

How well speech recognition really works

Character error rate on the FLEURS test split, measured 31 July 2026. 20% means one character in five is wrong. For reference, the same pipeline scores 8% on French. These numbers are not good, and we would rather you saw them than discovered them.

Eleven of these languages have no language code on any model we can serve. Whisper and Qwen3-ASR accept exactly the same eight African languages and reject exactly the same eleven, so you cannot tell either one it is listening to Igbo, and it will decide on its own that it is hearing English. Every African language in this table was auto-detected as something else.

Pick per language rather than by reputation: Qwen3-ASR is clearly better on Hausa, Igbo and Zulu, Whisper on Swahili. We do not route between them for you, because that would change your bill and your quality on a guess you never saw. Full method and the rest of the landscape in the write-up.

Cost by language

How many tokens the same sentence takes compared with English. 1.95x means a Swahili sentence costs almost twice as much as the identical English one.

Measured on FLORES-200, 1012 professionally translated sentences aligned across languages, run against each model's own tokenizer. Reported as the mean per-sentence ratio. A model reading 'not measured' is one whose tokenizer we could not obtain. We would rather say so than publish a number we have not run.

A cheaper tokenizer is not always a cheaper request

Models marked "after reasoning" think before they answer, and that thinking is billed even though it never reaches you. The same short Swahili reply costs a fraction of a cent on llama-3.1-8b and several hundred on kimi-k2.6, because the reasoning dwarfs the answer.

So a better multiplier above pays for itself on long answers and not on short ones. For one-line replies, a direct model is usually cheaper despite the worse tokenizer. Both numbers are here so you can work out which applies to your workload.

Reasoning models also need room. Send a small max_tokens and the budget goes on thinking, leaving you an empty answer you paid for, so we raise the ceiling for them. You are still only billed for the tokens actually produced.

Why this column exists

A price per million tokens assumes a token is worth the same in every language. It is not. The same sentence in Swahili or Amharic splits into more tokens than in English, so identical meaning costs more. A single flat rate hides that difference rather than removing it, and the rate above hides it too.

Some open models with the strongest African-language performance carry non-commercial licences. No provider can serve them to you, ourselves included, so they are absent from this list rather than swapped silently for a weaker model.

Every model here is open-weight

The weights are published, so you can download any model in this list, run it on your own hardware, and check the prices and language multipliers above against your own measurement. That is the difference between a number you are given and a number you can reproduce.