Summary
Two Minute Papers looks at DeepSeek’s latest release, V4.1 Flash 1017, which arrives only about five weeks after the previous 4.1 Flash. Despite the modest version bump, the video argues the jump in capability is large, particularly in building working software and operating a computer through the terminal.
The episode covers pricing and cost in detail. The API costs just over a dollar per million output tokens, roughly 15 times cheaper than Claude Opus 5.5. The host cautions that price per million tokens can be misleading, since in their own independent tests DeepSeek often used two to three times as many tokens as Claude Opus to finish the same task. Cost per task, not cost per token, is the better yardstick. The weights are also available to download, and the video discusses how open models are routinely distilled into smaller, cheaper systems that eventually run on consumer hardware.
The central question is how performance improved without any architecture change. Based on the wording of DeepSeek’s announcement and related research, the host suspects the key is learning from self feedback, where a model critiques and corrects its own work, such as spotting broken collision detection in a game it coded and fixing it. The video frames this as a possible shift away from the constant hunt for more training data.
📺 Source: Two Minute Papers · Published October 28, 2026
🏷️ Format: News Analysis
