Chinese company DeepSeek released an updated model, DeepSeek-V4.1-Flash-1017, on October 17. According to the developer, it runs faster, gives shorter answers, and in a number of tests on programming and AI agent work it beats significantly larger competitor models. The model's weights were immediately released as open access.
DeepSeek announced the release on social media X at 16:35 Beijing time. The new version already runs on the website, in the mobile app, and through the company's API, while the weights are published on Hugging Face and ModelScope. DeepSeek itself calls the update "minor," but the gains in tests are noticeable.
Developers don't need to change anything: the model is available through the API under the same name deepseek-flash, and prices haven't changed. Input tokens cost $0.3 per million (without cache), output tokens cost $1.2 per million during peak hours, and half as much off-peak. The context window remains 1 million tokens, and a response can contain up to 384 thousand tokens.
In DeepSeek's own tests, the new model improved on its predecessor's results in every category. In the DeepSWE v1.1 test, which checks the ability to fix real bugs in code, it scored 77.0 points versus 74.2 for the previous version, 74.0 for Claude Opus 5, and 73.0 for GPT-5.6 Sol. In Terminal-Bench 4.0, where the model works in the command line, the result rose from 31.2 to 42.3.
The main advantage is cost. According to DeepSeek, one task in DeepSWE costs on average less than $0.20, whereas the previous version cost about $0.50. Competing models on the same chart cost from $2 to more than $10 per task. At the same time, on difficult questions without access to tools (the HLE test), the model lags noticeably behind the leaders: 42.6 points versus 56.3 for Claude Opus 5.
For training, DeepSeek used a reinforcement learning method based on self-assessment (RLSF): the model uses its own confidence estimates and self-criticism as a reward signal, so neither human raters nor separate reward models are needed. The company says this helps the model create complete software projects from scratch — from games to 3D modeling.
Along with the model, DeepSeek for the first time opened its internal search engine to third-party developers through the Search API and Fetch API. According to the company, it already handles hundreds of millions of requests per day. Each of these services costs $0.75 per 1,000 calls.
It's worth bearing in mind that all the figures given are DeepSeek's internal tests, and there are no independent verifications yet. Nevertheless, the release once again shows how Chinese open models are putting pressure on closed Western competitors precisely through price.
Sources: The Decoder, Hugging Face (DeepSeek model card), DigIntoAI, Bloomberg, October 17, 2026.