AICodeKing revisited an earlier DeepSeek V4.1 Flash coding test after finding that thinking had been disabled in the first run. With maximum effort enabledReasoning effort is the amount of internal computational work an AI model applies before producing an answer or action., the creator's eight-task KingBench 3 score rose from 43 to 65 out of 80. This is a correction to the test setup, not evidence that the provider silently changed the model; the retest used a temporary pre-release API name and was not rerun against the final production name.
The transcript walks through a working elevator simulation, improved 3D contact-lens case and folding table, an SVG panda, an archery game, a math solver, a local fine-tuned panda-fact application and an interactive wristwatch. Several outputs improved materially, but the table still has intersecting parts, the archery result screen breaks after the targets are hit, and the panda-fact test data leaks training rows and includes inaccuracies. The source's score is a creator-run practical ratingEvaluation measures how well an AI system performs against defined tasks, criteria and failure conditions using repeatable evidence., not an independently standardized benchmark.
The creator estimates roughly 221 generated tokens per second from saved OpenCode response logs, including reasoning but excluding later tool execution. The video also relays DeepSeek's architecture and pricing claims; those are source-reported launch details rather than separately verified measurements. The useful takeaway is to check reasoning configuration, inspect saved outputs and test finished interactions before interpreting a coding-model comparison. Channel promotion, donation and subscription appeals are omitted.
Watch on YouTube
