DeepSeek V4 Pro beats GPT-5.5 Pro on precision

DeepSeek V4 Pro's rigid instruction following gives it a major edge for structured developer tasks, while GPT-5.5 Pro's tendency to improvise can break production applications.

• –DeepSeek V4 Pro won the coding task (python-log-redactor) by correctly combining regex patterns into a single replacer, avoiding potential ordering bugs that GPT-5.5 Pro's multi-regex solution introduced.
• –In vendor-delay-update, DeepSeek V4 Pro adhered closely to constraints without adding unnecessary handoff/escalation details or changing the recipient to a different department like GPT-5.5 Pro did.
• –GPT-5.5 Pro failed the structured schema constraints of the meeting-notes-summary task by introducing extra text to launch_date and formatting blocked_by as an array instead of a single value.
• –The models tied on the simpler messy-orders-to-json task, showing they are equally capable of standard data normalization tasks.