GPT Audio Mini: Speech In, Speech Out, at a Third of the Price

Most teams that add voice to a product start by chaining three models: speech-to-text, a text model, then text-to-speech. It works, but every hop adds latency and a place for meaning to get lost. GPT Audio Mini is OpenAI's cheaper option for collapsing that chain into one request: audio goes in, audio comes out, and the model reasons over the sound directly.

The catch is in where it runs. GPT Audio Mini is a Chat Completions model, not a Realtime one, and that single fact decides most of what it is good for.

What the model is

OpenAI describes gpt-audio-mini as "a cost-efficient version of GPT Audio". It accepts text and audio as input and can produce text and audio as output. The documented context window is 128,000 tokens with a maximum output of 16,384 tokens, and the knowledge cutoff is October 2023.

There are two dated snapshots on OpenAI's model page: gpt-audio-mini-2025-10-06 and gpt-audio-mini-2025-12-15, with the December one as the current default. The OpenRouter listing describes the newer snapshot as having an upgraded decoder for more natural-sounding voices and better voice consistency. That is the kind of change you hear rather than measure on a benchmark, which is a good reason to pin a snapshot rather than the bare alias once you have tuned prompts against one. Pinning model versions covers why an alias moving underneath you is the quiet failure.

OpenAI's model page lists function calling and prompt caching as supported features, and OpenRouter lists JSON mode and structured outputs as well. Tool calling on an audio model matters more than it sounds: a voice assistant that can look up an order status mid-utterance is a much more useful thing than one that can only talk.

The pricing, and where the money actually goes

Audio models bill two kinds of token at two different rates. On OpenAI's pricing page at the time of writing:

• GPT Audio Mini — audio $10 input / $20 output per million tokens; text $0.60 input / $2.40 output.
• GPT Audio — audio $32 input / $64 output per million tokens; text $2.50 input / $10 output.

On audio, the mini model is a little under a third of the full model's price in both directions. On text it is roughly a quarter. OpenRouter lists the text rates at the same $0.60 and $2.40.

The number to watch is audio output. A spoken reply is far more expensive per token than the same reply as text, so a design that returns audio only when the user is actually listening, and text everywhere else, can cut the bill substantially. Do not estimate audio token counts from a rule of thumb: run a representative sample of your own recordings through the API and read the usage object, which reports audio and text tokens separately. That gives you a real cost per minute of conversation for your speakers, your languages and your silence ratio.

Chat Completions only: what that rules out

OpenAI's model page lists exactly one supported endpoint for GPT Audio Mini: v1/chat/completions. It is not available on Realtime, Responses, Batch or Assistants, and it is not a dedicated transcription or speech-generation model either.

In practice that means each exchange is a bounded request. You send a complete audio clip, you wait, you get a complete response back. There is no interruption handling, no streaming microphone session and no server-side voice activity detection. For a live phone agent where the caller talks over the bot, that is the wrong shape, and OpenAI's own audio guide points that use case at the Realtime API instead.

The shape suits work that is naturally turn-based or offline:

• Voice notes in, a written or spoken summary out.
• Push-to-talk assistants where the user finishes speaking before the model answers.
• Generating spoken versions of short, dynamic text where you also want the model to rephrase it.
• Analysing tone or intent in recorded calls, where reading a transcript would lose information.

The request itself is ordinary Chat Completions with two additions: modalities set to include audio, and an audio object naming a voice and an output format. OpenAI's guide examples use voice: "alloy" and format: "wav".

Mini, full model, or a chained pipeline

There are three realistic choices for a voice feature, and the mini model wins only one of them.

Pick GPT Audio Mini when the reply needs to sound natural, the exchange is turn-based, and you are running enough volume that the price gap to the full model matters. The GPT Audio guide covers the larger sibling; the honest test between them is a blind listen on your own prompts, not a spec sheet.

Pick the full GPT Audio when the spoken answer carries harder reasoning, such as explaining a billing dispute or walking through steps that depend on each other. A smaller model is more likely to sound right and be wrong.

Pick a chained pipeline when you need the strongest possible reasoning, a model that is not audio-capable, or a full transcript as a first-class artefact for compliance. Chaining costs latency, but it lets you swap the middle model freely. It is also worth checking whether a newer audio model fits better: OpenAI's current audio guide uses gpt-audio-1.5 in its examples, so GPT Audio Mini is no longer the only option in the family.

How to evaluate it before you commit

Voice quality is subjective, but evaluating it does not have to be. A small, repeatable test is enough:

1. Record twenty real utterances from your users, including accents, background noise and people who trail off.
2. Run each through the snapshot you intend to ship, with your real system prompt and tools.
3. Score the content of each answer as you would a text model: correct, partially correct, wrong.
4. Have two people blind-rate the audio for naturalness and for whether the voice stays consistent across a multi-turn exchange.
5. Record wall-clock time from request to the first playable audio, and the audio token counts from the usage object.

That gives you three numbers per model: accuracy, perceived quality and cost per exchange. Best model for low latency covers why time to first output usually matters more to a listener than total generation time. If accuracy holds and the voices pass a blind listen, the mini model is the right default; if accuracy slips on the harder prompts, route those to the full model and keep the mini one for everything else.