Beyond text
Modern prompts often include images, audio, or files. The mechanics differ sharply by provider.
Vision
- Anthropic — image as a content block (base64 or URL); supported on most Claude models.
- OpenAI — image_url content type; vision is part of the multimodal models (gpt-5.5).
- Gemini — inline_data with mime_type; native multimodal across the family.
Audio
- OpenAI — Whisper for transcription; Realtime API for live voice.
- Gemini — native audio understanding inline.
- Anthropic — text-only on most public models in 2026; pair with separate transcription.
Files
- Anthropic — Files API (upload once, reference by id).
- OpenAI — Files API + Assistants attachments.
- Gemini — Files API for large media.
Implications for prompts
Multimodal inputs change what the prompt can refer to ("the second image," "the chart in figure 3"). Reference content explicitly. Vision tokens cost differently from text tokens; track them separately.