LLMs, GenAI, RAG, agents, evaluation · insight-1518
Downsample images to the resolution the task needs, because visual tokens scale with pixels
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- ML engineer, Applied AI
- Industries
- aviation, insurance
Summary: situation, action, reason, boundary
Situation
A baggage-claims assistant sends full-resolution phone photos and scanned forms with every request, and input cost and latency run well above the text-only estimate.
Action
Read image tokens from usage on a sample, set a target long edge per task type (small print needs more than a damage photo), resize in code before sending, compare accuracy against full resolution on a labelled set, and reference images by file id in long chats.
Reason
Anthropic documents that an image costs about ceil(width/28) times ceil(height/28) visual tokens up to a per-tier cap, and that a high-resolution tier can use up to roughly three times more visual tokens than the standard tier for the same image.
Boundary
Multimodal assistants over photos, screenshots and scanned documents.
Symptom or risk
Image tokens dominate the bill, multi-turn chats resend them each turn, and nobody knows which images needed the resolution.
Exceptions
Dense documents and fine print, where downsampling removes legibility and the full resolution earns its cost.
Trade-offs
Preprocessing code and a resolution policy per task, against a lower and more predictable token bill.
Pitfalls
Several rounds of lossy compression, which the documentation warns can make text unreadable even when the pixel count is right.
Lessons and concepts
Claims and their sources
Anthropic's vision documentation states that an image costs ceil(width/28) x ceil(height/28) visual tokens, that images beyond a per-tier long-edge or token limit are downscaled, that high-resolution models can use up to roughly three times more visual tokens than standard ones, and that pre-resizing controls cost (as read on 2026-10-12).