I Gave DeepSeek a Token Limit. It Ignored Me.
A hands-on test of V4-Pro's default reasoning mode ⚠️ This article was generated with AI assistance (Claude). Every number comes from real calls the author made to DeepSeek's official API on the day of testing, and sample sizes are included with each conclusion. The author is an independent developer with no affiliation to DeepSeek or OpenRouter. Tested on 2026-08-14 using deepseek-v4-pro on…
A test was conducted on DeepSeek's V4-Pro model, focusing on its default reasoning mode. The article reveals that reasoning is activated automatically without any specific parameter being set. During the testing, a single Liar's Dice decision consumed an enormous 3,072 output tokens, all of which were used for reasoning with no visible answer provided.
The OpenAI-style fix, where a max_completion_tokens parameter was set to 3072, made no difference. The API returned an HTTP 200 status code, generated 15,809 tokens over 222 seconds, and behaved as if it had never seen the parameter. Interestingly, even a made-up parameter like totally_bogus_param: true was accepted by the API. This means that hidden parameters are swallowed, and you cannot determine if a limit takes effect until the bill arrives.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
