I've previously found general AI models fairly unusable at generating retro assembly programs, but the recently released Claude 4 looks to have made the jump into useful territory.

Claude Sonnet models' scoring in percentage over time in my semi-public DOS FASM benchmark, easy/medium difficulty for someone with some experience in assembly:

Token price has remained the same. Zero-shot, no test-time thinking.

Sample output for the tasks 'display information about mouse input' and 'load and display a paletted image':
https://leikareipa.github.io/dosbox/#/ai-asm- … 4-sonnet/mouse/
https://leikareipa.github.io/dosbox/#/ai-asm- … -sonnet/parrot/