PixelPinch

Built-in macOS screenshots are huge, and vision-capable LLMs charge you image tokens for every pixel. PixelPinch sits in your menu bar and rebinds the standard screenshot keys to a smarter pipeline: capture, downsize, JPEG-compress, and write straight to the clipboard.

The result is a screenshot that looks identical to a human at chat resolution, but costs roughly 4× fewer tokens to send to an AI. No round trip through Preview, no manual export, no third-party services.

Vision models don't bill images by megabytes — they bill by pixels. Claude, for example, slices each image into 28×28-pixel patches and charges roughly one token per patch: tokens ≈ ⌈w/28⌉ × ⌈h/28⌉.

Mac screenshots are captured at Retina 2× scale. A 700×500-point window lands as a 1400×1000-pixel image: 50 × 36 patches ≈ 1,800 tokens. PixelPinch downsizes it back to 1× — 700×500 pixels, 25 × 18 patches ≈ 450 tokens. Half the width times half the height is a quarter of the patches: ~4× fewer tokens, and at chat resolution the two look the same to both you and the model.

The asterisk: the numbers in the animation are that worked example, not a guarantee. Every provider counts differently (ChatGPT bills 512-px tiles, Gemini 768-px tiles), most cap what a single huge image can cost (many Claude models cap around ~1,600 tokens per image; newer high-res Opus models around ~4,800), and JPEG compression shrinks upload size rather than token count. The honest summary is “roughly 4×, varies by model and window size” — which is why we say roughly.