DeepSeek's V4-Flash aims at agents at a very low price

DeepSeek has released DeepSeek-V4-Flash-0731, a 304 billion parameter model aimed at agentic work and priced aggressively. On Simon Willison's reading of the numbers, it runs at about $0.14 per million input tokens and $0.27 per million output, which works out to roughly $0.028 for a typical task on Artificial Analysis's benchmark while scoring around 50 on their intelligence index. That puts it ahead of MiniMax M3, a 428 billion parameter competitor, at a fraction of the active compute.

DeepSeek's model card leads with agentic benchmarks: 82.7 on Terminal Bench 2.1, 76.7 on Cybergym, and 68.7 on DSBench-FullStack, among others. The company says the release beats its own larger V4-Pro preview despite a much smaller number of active parameters. The model ships under the MIT license, uses a speculative decoding module for faster inference, and exposes a reasoning_effort control with low, high, and max settings. DeepSeek recommends allowing up to 384K output tokens at the higher levels, which points at long agent runs rather than short chat replies.

Why it matters

If you build agents and judge models by cost per finished task rather than headline size, V4-Flash is a cheap, MIT-licensed option worth running through your own eval before you commit to a pricier closed model.