An independent developer testing DeepSeek's V4-Pro model discovered that the API ignored specified token limits, leading to excessive token consumption and increased costs. The model's default reasoning mode generated thousands of tokens for internal thought processes, resulting in empty visible answers and inflated bills. While parameters like `max_completion_tokens` and even custom ones were accepted by the API without error, they did not appear to enforce any actual limits. The developer found that explicitly disabling reasoning was necessary to achieve predictable costs and performance for their application, Kai. AI
IMPACT Highlights potential issues with API parameter enforcement and cost management for developers integrating LLMs.
RANK_REASON Developer testing of an existing model's API parameters and behavior.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
