Why does attaching a screenshot make your assistant noticeably slower to start answering?

When you attach a screenshot, two timers start that did not exist for a text-only message. **Timer 1: transport and preprocessing.** The image (often base64-encoded, so 33% larger than the raw bytes) uploads over HTTPS. The provider then decodes it, resizes to its internal tile dimensions, and tokenises into vision tokens. All of this happens before the model sees a single token. For a multi-megabyte screenshot on a slow connection, transport alone can add hundreds of milliseconds. **Timer 2: prefill grows.** A single image typically adds 500 to 2000 vision tokens to the prompt depending on resolution and detail mode. The model must process the entire prompt (text plus image tokens) before it can emit the first output token. Prefill time scales with token count, so time to first token stretches noticeably.

**Two cost centres** show up between the user pressing send and the assistant streaming the first word. **Transport and preprocessing.** The image is encoded (usually base64, inflating bytes by 33%), uploaded over HTTPS, decoded by the provider, resized to internal tile dimensions, and tokenised. On a slow upload link this is a real chunk of wall-clock time, often 200 to 800 ms for a multi-megabyte screenshot. Provider-side preprocessing adds another 100 to 300 ms before inference even starts. **Prefill on a longer prompt.** The model must process every token in the prompt before sampling the first output token. A single image at high detail can add 1500 to 2500 vision tokens; multiple images compound. Prefill time on modern frontier models is dominated by attention-over-prompt cost, which scales near-linearly in token count for short contexts and worse for long ones. The image easily doubles or triples prefill time versus a text-only message. **Mitigations, mapped to each cost:** For transport + preprocessing: - Compress dimensions before upload (long edge ~1280 px for general use). - Use prompt caching on repeated images so the second turn skips preprocessing entirely. - Skip base64 where the SDK supports binary upload. For prefill: - Crop to the relevant region when the question is local. - Use low detail mode for gist questions (caps vision tokens to about 85). - Cache the image prefix so subsequent turns pay only the cached-prefill cost.