DeepSeek updates V4.1 Flash with better agent scores and shorter answers

DeepSeek released V4.1-Flash-1017, a minor update to V4.1 Flash, on its website, app and API, with open weights under the MIT licence on Hugging Face. DeepSeek's own tests show gains over the previous version on every benchmark it lists, for example Terminal-Bench 4.0 up from 31.2 to 42.3, and it says the model finishes tasks faster with shorter outputs, so jobs cost less. It keeps a 1M token context window, and the weights are 763B parameters, so self-hosting needs serious hardware. DeepSeek also opened its Search and Fetch APIs to everyone at $0.75 per 1,000 calls per endpoint; calls made through DeepSeek Harness, Claude Code and Codex stay free for DeepSeek API users until billing starts on 31 October. On Hacker News (673 points, about 575 comments), the top comments split. Several daily users say speed, low price and visible reasoning make it feel better than the benchmarks suggest, and some describe letting DeepSeek do the coding while Opus or Sol plans and reviews. One commenter who measured it found about 20% fewer tokens at each reasoning effort with the same results. Others push back: in one hands-on code review test, GPT-6.1 Sol found more relevant issues in about 4.5 minutes than the new DeepSeek did in 20. No consensus on whether it matches the frontier models. Sorted take: the benchmark numbers are DeepSeek's own. If you run agents on a budget, try it on one real task next to your current model and compare the cost per finished task, not just the score.