DeepSeek V4 Pro Outperforms GPT-5.5 Pro in Precision

In a comparison of precision and instruction-following ability between DeepSeek V4 Pro and GPT-5.5 Pro using 4 freshly generated text tasks, DeepSeek V4 Pro scored 38.0 points, outperforming GPT-5.5 Pro (33.0 points). The grading for this test was performed by the grok-4-1-fast-non-reasoning model.

The key test results are as follows:

• python-log-redactor (code writing): DeepSeek won by accurately handling nested patterns with a single regex. GPT-5.5 Pro showed regex ordering errors and matching flaws.
• vendor-delay-update (business message writing): DeepSeek accurately followed the prompt's requirements and tone. GPT-5.5 Pro deviated somewhat from the instructions by adding unrequested details.
• meeting-notes-summary (summarization and JSON generation): DeepSeek won by perfectly complying with the specified JSON schema. GPT-5.5 Pro violated the structure due to data type errors and the inclusion of conditional text.
• messy-orders-to-json (data normalization): Both models produced valid JSON, resulting in a tie.

In conclusion, DeepSeek V4 Pro was evaluated as a more disciplined, accurate, and reliable model in precision tasks, where even small deviations directly lead to actual failures.