Reducing Image Context Cost in Claude Code — Thinking in Pixel Dimensions and Cutting with Downscale, Crop, and Text Substitution

When you paste screenshots into Claude Code over and over, the conversation context grows faster than expected. The cause is the number of pixels, not the size of the file. This post walks through the token formula and long-edge caps documented by Anthropic, and then orders the practical options (downscaling, cropping, substituting text or structured data, and how you handle history) by how much they actually help.

Reducing Image Context Cost in Codex — Cutting Tokens with 32×32 Patches and detail: low / high / original on GPT-6.1, GPT-6, and GPT-5.6 Models

When you keep pasting screenshots into Codex, the context runs down faster than expected. Among the GPT-6.1, GPT-6, and GPT-5.6 models Codex ships with as of November 2026, the ones in OpenAI's official guide (GPT-6-Astra and the GPT-5.6 family) use the 32×32 px patch method with a `detail` parameter and a 30,000-patch per-image cap. This post walks through the math, then orders the practical options (detail switching, downscaling, cropping, and text substitution) by how much they actually help.