Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs
decoder ·
Chinese AI company Deepseek has released an updated version of its Flash model, named V4.1-Flash-1017, which demonstrates significant improvements in agentic and coding tasks. The model utilizes reinforcement learning with self-feedback to enhance end-to-end software development capabilities while producing shorter outputs that drastically reduce costs. In internal benchmarks, this smaller model outperforms larger competitors such as Claude Opus 5 and GPT 5.6-Sol in accuracy and efficiency. Specifically, the update achieves high scores at a fraction of the cost per task compared to previous versions and rival offerings. Alongside the model release, Deepseek is opening its internal search engine to developers via new Search and Fetch APIs. These API endpoints are priced at $0.75 per 1,000 calls, with free access available for certain tools until late October 2026. Meanwhile, the article notes potential regulatory challenges in China regarding the publication of open-weight models.
decoder