DeepSeek V3.1: A New Chapter for Smarter AI Agents

DeepSeek has officially unveiled V3.1, a major update that the team describes as their first real step into the “era of agents.” The model was quietly available for two days before the formal announcement, and benchmarks already show why the upgrade matters. On tasks like SWE-bench, which tests how well an AI can navigate and solve problems in code repositories, and Terminal Bench, which measures command-line performance, V3.1 demonstrates clear gains.

What makes the release especially notable is the unification of speed and reasoning. Unlike its predecessor R1, which required a separate branch for deep reasoning, V3.1 can smoothly switch between quick responses and more deliberate, step-by-step thinking. The model also achieves this efficiency while using 28 to 46 percent fewer tokens for reasoning, which translates into faster answers and lower costs.

Pricing and Accessibility

Starting September 5, new pricing takes effect on the API. One million output tokens will cost $1.68, while input tokens range from $0.07 with caching to $0.56 without it. For developers who rely on fast responses, this is a slight increase. For those who need reasoning-heavy tasks, however, the reduced token usage makes the model significantly more affordable. The only drawback is the removal of the popular nighttime discount.

For casual users, the news is even better: V3.1 is completely free on the web and in mobile apps. This balance of power, efficiency, and accessibility signals an exciting step forward, not just for DeepSeek but for the growing world of AI agents.