Codex "Audio input is not available": why the chat locks on GPT-6.2 Sol
Back to the blog
Articlecodexopenaitroubleshootingdev toolsverboo code

Codex "Audio input is not available": why the chat locks on GPT-6.2 Sol

MafraNovember 22, 20266 min read

The chat was working. A tool returned a chunk of audio, the next turn came back with HTTP 400, and from then on nothing got through. Not plain text, not even a one-word "hi". The error is always the same, and it says nothing about audio in your code: Audio input is not available.

It is not your project. The audio item was saved into the conversation history and comes back on every request, even when you type text. As long as it sits there and the model declares that it accepts audio, the client does not strip it, the backend rejects the whole payload, and the turn dies before any command runs.

This article shows the line of code behind the behavior, the command that puts the chat back to work in seconds, and what to do so the class of error does not come back.

Why does Codex refuse even a text message after an audio clip?

Because the history is resent in full on every turn, and the media cleanup does not happen when the model advertises support for that modality. The payload goes out with the old audio inside it, the endpoint rejects the entire request, and no part of it is processed.

This is the response you get, with the param field pointing at the whole input rather than at a specific attachment:

HTTP 400 Bad Request

{
  "error": {
    "type": "invalid_request_error",
    "param": "input",
    "message": "Audio input is not available."
  }
}

Keep the detail that confuses everyone: the error does not belong to the turn where the audio came in. The turn that produced the audio succeeds. The one that breaks is the next one, and every one after it, because the item is already written into the session.

Which line of code explains this?

The one in strip_audio_when_unsupported, inside the client context normalizer. It exists to remove media the model cannot accept, and the first thing it does is return when the model declares the modality:

// codex-rs/core/src/context_manager/normalize.rs
pub(crate) fn strip_audio_when_unsupported(
    input_modalities: &[InputModality],
    items: &mut [ResponseItemEnvelope],
) {
    if input_modalities.contains(&InputModality::Audio) {
        return;   // model declares audio: nothing is removed
    }
    // ... walks messages and tool outputs and replaces
    // InputAudio with a text placeholder
}

The reasoning is sound. If the model accepts audio, there is no reason to delete audio. The problem sits at the input of that decision: the model catalog advertises audio for a model whose endpoint actually refuses it. The guard never fires, and the cleanup that would have saved the conversation stays off.

Here is the side by side, measured on 2026-11-19 at 18:30 UTC in an authenticated catalog fetch on version 0.170.0:

ModelDeclared input_modalitiesWhat happens to audio in history
gpt-6.2-soltext, image, audioIt is not removed. The payload is rejected with 400.
gpt-6.1-soltext, imageIt is replaced with a text placeholder. The turn goes through.

The symptom also shows up for people who never asked for audio. All it takes is a tool that renders a preview, an ffmpeg call that extracts an MP3, a transcription step that returns the file alongside the result. What matters is not the intent of the turn, it is the format of the tool output.

How do I recover the same chat right now?

By switching models inside the chat itself. The history does not need to be deleted, edited, or exported, and the earlier work stays where it is:

  1. Type /model.
  2. Pick gpt-6.1-sol.
  3. Resend the text message that failed.

With the model that declares only text and image, the cleanup runs, the audio leaves the payload, and the turn completes. The audio item stays in the saved history, it just is not sent anymore. This was reproduced across three different environments in the public report of the problem, with a consistent result: the conversation answers again, without a fork and without touching the session files.

Two honest caveats. First: you recover the conversation, not the audio. The model does not process the original clip, it only follows the thread of text. Second: if the task depends on hearing the file, switching models clears the block but not the task. In that case the path is to redo the step with a tool that returns a file path instead of inline audio.

Why does a new chat work while "Branch in new chat" does not?

Because a new chat is born without the audio item in its history, and a branch copies the whole history, including the item that is causing the 400. One report on the issue reproduced exactly that: the branch came back with the same error, and only a blank chat continued.

That is community diagnosis, not documented behavior: the report did not capture the branch payload, so it is not possible to state whether the audio was resent or whether another item broke along with it. Treat it as a lead, not a rule. The recovery with payload verification behind it is the model switch in the same chat.

How do I keep the error from coming back?

By treating audio as tool output data, not as a detail. Three decisions matter:

DecisionWhy
A tool that produces audio returns the file path, not inline audioThe item never enters history, so no model can resend it.
Media previews stay out of tool outputAn input_audio preview poisons the whole session, not just that turn.
An audio-capable model is used only when the flow truly needs audioWith no audio declared, the automatic cleanup works in your favor.

If you maintain an agent that talks to this client, the practical rule is one: never return inline media in a tool result when the content can be referenced. A file path, a signed URL, an object id. All of them survive a model switch. The inline blob does not.

What is still unresolved?

The issue remains open, with no published fix, and the catalog still advertises audio for gpt-6.2-sol. Until both ends agree, the behavior is this: the client trusts the declaration, the endpoint refuses, and the user is left with a stuck chat that answers 400 to anything.

A community patch exists that tries to work around it in the client, but it compares the error message literally. That covers Audio input is not available. and leaves out the Audio input is not supported for this request. variant, which arrives with a different error code. Until the fix comes from above, the reliable path is still to keep inline audio out of the history.

What locks the conversation is not the audio itself, it is having nowhere to switch when a model refuses the payload. In Verboo Code, /model swaps the engine mid-session with the history intact, so the same maneuver that recovers a Codex chat is the normal operation of the day: if a model refuses what you sent, you change models and continue from where you stopped. Open Verboo Code and try a model switch mid-session.

Enjoyed this article?
Share knowledge with your network.
// Read also

Related articles