Of the 291 models available on OpenRouter right now, 22 accept audio input. That is under 8%. Every model announcement for two years has used the word multimodal, and for almost all of them it means text and pictures. If you have audio, your choice of model collapses to a shortlist.
The actual state of multimodal
Counting input modalities across the current catalogue: text is universal, images are common, files are widespread, video is uncommon, and audio is rare. Fifteen models accept all five, and two thirds of those are Google's Gemini line.
Meta's Muse Spark 1.2, released yesterday at $1.25 per million input tokens and $4.25 output, takes all five and pairs them with a 943,718-token output ceiling and a 1M context window. Thinking Machines' Inkling Small, from 30 July, takes text, image and audio at $0.45 and $1.20.
The cheapest way to get all five modalities today is Gemini 3.5 Flash Lite at $0.30 and $2.50. That is a real option and not an obvious one, because it is a small model that nobody discusses as a multimodal flagship.
Why this catches people out
The word does no work. A model that reads images and a model that ingests an hour of meeting audio are both described as multimodal, and the gap between them is the entire feature.
The practical consequence is that a plan formed from announcements can fail at implementation. You choose a model on quality and price, then discover it will not take the input your product is built around, and your shortlist was never the shortlist you thought it was.
The reverse error is also common and more expensive: assuming audio requires a frontier model. It does not. The full-modality list includes several small, cheap models, and if your task is transcription or extraction rather than deep reasoning about the content, one of those is likely sufficient.
Audio is not one capability
Worth separating three things that get bundled together.
Transcription is turning speech into text. Dedicated speech models generally do this better and far more cheaply than general models with audio input. If transcription is all you need, a general model is the wrong tool.
Understanding is answering questions about what was said, including tone, who spoke, and what was implied. This is where audio input on a general model earns its place, because the model reasons over the audio rather than over a lossy transcript.
Generation is producing speech. That is a separate capability entirely and is not what an audio input modality gives you.
Most people asking for "an audio model" want the first, occasionally the second, and are quoted prices for the second when the first would have done.
How to choose when the shortlist is short
When only fifteen models can take your input at all, the usual advice about testing widely stops applying. The question becomes narrower: of the handful that can do this, which is best, and is the cheapest one good enough?
That is a comparison worth doing carefully precisely because the field is small. Run the same audio through Gemini 3.5 Flash Lite, Muse Spark and whatever else qualifies, and look at the outputs together. Three candidates is a manageable test, and the price spread between them is over four times.
The general lesson holds: capability lists tell you what is possible, not what is good. The only way to know whether the $0.30 model handles your audio as well as the $1.25 one is to put the same file through both.
Frequently Asked Questions
How do I check a model's input modalities?
The OpenRouter model listing publishes an architecture object with an explicit input modality array for each model. Read that rather than the launch announcement, since the announcement is written to sound broad and the array is the thing the endpoint enforces.
Should I use a general model or a dedicated speech model?
For transcription, a dedicated speech model, almost always: better accuracy and far cheaper. For reasoning about audio content, including tone and speaker dynamics, a general model with audio input, because transcribing first throws away exactly the information you are asking about.
Is audio input going to become standard?
We would expect it to spread, but we are not going to predict a timeline. What we can say is that as of early August 2026 it is present in under 8% of models, and that building on the assumption it will arrive in a particular model is a bet rather than a plan.
