[BUG] Read tool doesn't support audio files but model claims to perceive audio tonality/emotion

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Code's Read tool does not support audio files (e.g., .mp3, .wav, .m4a), yet the model will claim to perceive qualities like tonality, emotion, and mood when asked to analyze an audio file. This creates a misleading experience where Claude confidently describes audio characteristics it has no ability to process.

For example, when you use the Read tool on an audio file and ask Claude to describe it, the model may respond with detailed commentary about the "warm tone" or "upbeat energy" of the audio, when in reality it has no access to the audio data at all. The Read tool either returns an error or garbled text, but Claude fabricates a plausible-sounding analysis anyway.

The inconsistency between the stated/implied capability and the actual tool limitation is the core issue.

What Should Happen?

One of two things should happen:

1. Support audio files in the Read tool: The Read tool should be able to ingest audio files and pass them to the model for genuine multimodal audio analysis (similar to how it handles images), OR

1. Clearly refuse and explain: When a user asks Claude to analyze an audio file, the model should clearly state that the Read tool does not support audio files and that it cannot perceive or analyze audio content, rather than confabulating a response about the audio's qualities.

Either path would resolve the inconsistency. The current behavior where Claude fabricates audio analysis is worse than an honest refusal.

Error Messages/Logs

No specific error output -- the Read tool may return garbled binary data or an error message depending on the audio file format, but the model proceeds to fabricate an analysis regardless.

Steps to Reproduce

1. Open Claude Code in a terminal
2. Have any audio file available (e.g., song.mp3, recording.wav)
3. Ask Claude to analyze the audio file using the Read tool, e.g.: "Read song.mp3 and tell me about the mood and tonality of this audio"
4. Observe that Claude produces a detailed analysis of the audio's emotional qualities, tone, and mood despite having no actual ability to process audio data
5. Note that the Read tool either errors out or returns meaningless binary data for audio files, yet the model responds as though it successfully perceived the audio

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

Latest (as of March 2026)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

This issue affects user trust. When a tool confidently produces analysis of data it cannot actually process, it undermines confidence in all of Claude Code's outputs. Users who rely on Claude's stated capabilities may not realize the audio analysis is fabricated, leading to incorrect conclusions or wasted time.

The Read tool documentation states it can read images (multimodal), but makes no mention of audio support. The model should be aware of this limitation and communicate it clearly rather than hallucinating audio perception.