ChinaDeepSeek cut cost per coding task 60% to 1.3 yuan.

China’s DeepSeek says its open-weight model beats GPT-6 Astra and Claude Opus 5 on DeepSWE.

The V4.1-Flash-1017 model passes 77.0% of DeepSWE tasks, 2.8 points above its predecessor, and its weights are downloadable on HuggingFace.

These benchmark results come from DeepSeek, not an independent lab.

DeepSeek begins billing search and fetch calls from coding tools at 00:00 Beijing time on 1 November.

How each outlet framed it
Zhidx leans favorable
cites DeepSeek's cost cut from about 3.3 to 1.3 yuan per task and the 244 tokens/second speed, plus the 5-hour outage refund

Sources: Zhidx