Image attachments inflate prefill (hundreds to thousands of vision tokens to process) and add transport plus preprocessing overhead before inference starts. Downscale, use prompt caching, and crop to attack both.
Imagine you ask a friend a question by text. They reply quickly. Now you also send them a big photo of your screen and ask the same question. They have to wait for the photo to download over slow Wi-Fi and then study it before they can answer. Two separate slowdowns: the photo took time to arrive, and reading it took longer than reading your short text. Sending a smaller crop of just the part you care about helps both. If you keep sending the same photo, remembering it from last time helps even more.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Time to first token (TTFT) is the latency users actually feel. It is the gap between pressing send and seeing the first character of the response stream in. For text-only chat with a modern frontier model in 2026, TTFT lands in the 200 to 400 ms range and feels instantaneous. Attach a screenshot and the same model jumps to 700 to 1500 ms, which feels noticeably sluggish.
The slowdown is not the model 'thinking harder'. It is the sum of two cost centres that did not exist in the text-only path: transport plus preprocessing of the image, and prefill expansion from the vision tokens the image produces. Both happen before the model generates a single output token. Understanding where the time goes is the prerequisite for fixing it.
This walkthrough breaks down the latency budget, explains why each cost exists, and lays out the mitigations in order of leverage.
Mental model: every image attachment buys you two extra timers, both before the first output token. Optimise each separately or accept the latency hit.
The latency budget, line by line
A typical text-only chat request to a frontier model in 2026 looks like:
- Network setup and TLS: ~30 to 50 ms
- Request body upload: ~10 to 30 ms
- Server-side queueing and routing: ~10 to 50 ms
- Prefill on the prompt (a few hundred tokens): ~100 to 200 ms
- First decode step: ~20 to 40 ms
- TTFT total: ~200 to 350 ms.
Add a 2 MB screenshot at default resolution:
- Network setup and TLS: ~30 to 50 ms
- Base64 encoding inflates payload to ~2.7 MB; upload over typical broadband: ~200 to 400 ms (much worse on mobile)
- Server-side queueing: ~10 to 50 ms
- Provider-side image preprocessing (decode, resize, tile, tokenise): ~100 to 300 ms
- Prefill on prompt + 1500 to 2500 vision tokens: ~300 to 600 ms
- First decode step: ~20 to 40 ms
- TTFT total: ~700 to 1500 ms.
The diff comes from two places: ~300 to 700 ms of transport plus preprocessing that did not exist before, and ~200 to 400 ms of extra prefill from the vision tokens. Both regions are addressable; both regions live entirely before the model's first output token, which is why generation parameters (temperature, max_tokens) cannot help.
One subtle point: queueing is shared between the two paths but can spike when the provider is under load. Image-containing requests sometimes route to different inference pools with different queue depths. Cold model loads on bursty traffic patterns can add hundreds of milliseconds that look like 'image latency' but are actually pool-allocation latency.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Claude Opus 4.7 and Sonnet 4.6 vision both show TTFT lifts of 300 to 800 ms per high-resolution image attachment.
- OpenAI GPT-5.5 vision exposes a low detail mode that caps image tokens to a small fixed budget for fast scene-gist tasks.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow much does base64 actually add versus binary upload?
33% size inflation plus encode/decode CPU; on fast connections the bandwidth cost is small but on mobile 4G it can add 100 to 300 ms; binary upload paths skip both.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Blaming inference. The latency hit is mostly prefill and preprocessing, both of which happen before the model emits a single output token. Tuning generation parameters does not help here.
60 second bullets to scan on the way to the call.
The two distinct latency cost centres for image attachments
What prefill is and why it stretches with token count
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.