Test: sending text as images cuts Claude tokens by up to 81% but breaks exact values

An independent test of 144 calls to Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1 found that rendering context as PNG images cuts token usage by 64–81%, but models struggle to reproduce numbers and identifiers. Opus 5.5 did best, recovering 58 of 60 IDs and 44 of 60 numbers.

• Token savings ranged from 64% for English prose to 78% for a server log
• Opus 5.5 got 44 of 60 numbers right, but only 3 of 10 in Korean text
• IDs came back from pxpipe's text factsheet, not from the image
• Models sometimes returned plausible wrong values instead of null