DeepSeek has released V4.1 Flash 1017 on its website, app and API. The company describes the update as an improvement in benchmark performance, task completion speed and output length over the previous version. Those comparisons are vendor-reported results.
What the coding update targets
The announcement focuses on end-to-end project development, including games and three-dimensional scenes. DeepSeek says reinforcement learning with self-feedback supports this work and shows several project demonstrations.
It also identifies DeepSeek Harness, Tencent CodeBuddy and WorkBuddy as optimisation targets. Users of DeepSeek Harness are advised to upgrade to v0.2.1-rc.1 or later. Open model releases are linked through Hugging Face and ModelScope.
A project demonstration shows the kind of output the vendor wants to support. It does not establish that a model will complete an unfamiliar repository correctly, preserve its tests or meet a particular team’s review standards.
Search and Fetch have separate call prices
DeepSeek also says its Search and Fetch APIs are open to everyone. Each endpoint lists a price of $0.75 per 1,000 calls.
There is a temporary exception for calls made through server-side tools in DeepSeek Harness, Claude Code and Codex by DeepSeek API users. The announcement says billing for those calls starts on October 31, 2026, at 16:00 UTC. That is a dated free period, not a promise that these tools will remain free indefinitely.
What to verify in an existing workflow
Franklin’s practical interpretation is to assess the model update and the retrieval bill separately. A useful local comparison would keep the repository, task and checks consistent, then record completion quality and the retrieval calls needed to reach it. Shorter model output alone does not measure whether the delivered project works.
Teams trying the server-side tools should also record the billing transition in their own cost assumptions. Read DeepSeek’s release note for the version recommendation, endpoint prices and exact transition time.