What's this about?
simonw ran his pelican-on-a-bicycle SVG test on DeepSeek-V4-Flash-0731 and got a disappointing result — until he raised the reasoning setting from default to "high," which produced a much better drawing.
What exactly is "reasoning effort"?
Modern models can spend a variable amount of hidden "thinking" tokens before answering — like choosing how long to work on a draft before handing it in. Providers expose this as a dial (low/medium/high). More effort costs more money and time but often yields better answers.
Why it matters
Model comparisons are becoming meaningless without their settings: the same model can look mediocre or great depending on one dial. When you see a benchmark or a demo, ask what effort level it ran at.