How much screenshot is enough?

A larger screenshot preserves more detail, but it also costs more to transport and process. The useful question is not whether bigger is better. It is which detail survives at the budget the receiving model can actually use.

Detail helps, but not in a straight line

Wang et al.'s Qwen2-VL study sweeps the same model across rising image-token budgets and watches accuracy move¹¹Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution Wang et al., 2024. On InfographicVQA, reading-heavy accuracy climbs steeply with detail:

InfographicVQA improves sharply as detail rises. Fine-OCR peaks earlier, then declines. The same model and image budget do not produce one universal curve across tasks. Resolution is a choice with an objective, not a quality dial that should always be turned to maximum.

The receiving model sets part of the budget

Vendor handling differs and changes over time. Anthropic currently documents standard and high-resolution model tiers with different long-edge and visual-token limits²²Claude vision documentation (image sizing & token cost) Anthropic, 2026. OpenAI documents several paths: low, high, original, and automatic detail, with patch- or tile-based accounting depending on the model³³Images and vision: tile-based tokenization OpenAI, 2026. These are useful constraints, not a timeless universal constant.

That variation is the reason a product should not hide the image it actually served. The dimensions matter because the model reasons over the delivered rendering, not the local original.

Too small fails differently

On VBench, which asks questions about small details in high-resolution images, the tested models struggled when the relevant detail was lost in static downsampling⁴⁴V: Guided Visual Search as a Core Mechanism in Multimodal LLMs Wu & Xie, 2023. A small payload can be efficient and still fail the task it was meant to support.

The Noru decision

Noru preserves the full-resolution original locally, then serves a capped rendering with its delivered dimensions declared in the evidence packet. That keeps the source available while making the agent-facing trade-off visible.

There is no universal ideal pixel count. Preserve the original, cap the served image deliberately, and disclose the delivered dimensions. Detail and cost should be a visible trade-off, not a hidden assumption.