Michael JohnsonOn October 17, 2026, DeepSeek updated the V4.1 Flash model to Flash-1017 and announced open access to its Search & Fetch API. The updated model is available on the website, in the app, and through the API; its weights have been published on Hugging Face and ModelScope. The practical takeaway for developers is straightforward: check which version your existing call receives, and account separately for future charges for search requests.
According to DeepSeek’s October 17 announcement, API calls still use the name deepseek-flash, and neither the calling method nor the model price has changed. Calls to that name are now served by the updated version. That means the version is less firmly pinned by the call name than it might appear: for repeatable quality evaluations, teams should record the test date, prompts, and results, then recheck critical scenarios after the update.
What changed
DeepSeek presents Flash-1017 as an improvement in task execution speed, response length, and benchmark results. The company links the release to end-to-end project development tasks and provides examples of creating games and 3D scenes. These materials describe the optimization goals and demonstrations; on their own, they do not establish the expected quality for a particular repository or workflow.
The published release includes comparative results for several benchmarks, including DeepSWE v1.1 and Terminal-Bench 4.0. To interpret the figures, it is important to read the conditions for each test: DeepSeek separately specifies the harness environments and scoring modes in its API changelog. Teams choosing a model for a coding agent would be better served by rerunning several of their own tasks with the same tool limits, budget, and acceptance criteria than by directly using an aggregate ranking to predict performance.
Flash-1017 weights are available under the MIT license on the Hugging Face model page. This makes it possible to explore self-hosting; the availability of weights alone does not indicate what compute resources a particular team will need. First check the project configuration and instructions, then evaluate a deployment on the target infrastructure.
API search: the free period has an end date
On the same day, DeepSeek announced that its Search & Fetch API was open to everyone. The announcement lists a price of $0.75 per 1,000 calls for each endpoint. Calls to these tools through the DeepSeek Harness server-side tools, Claude Code, and Codex are currently free for DeepSeek API users; the company set the start of billing for October 31, 2026, at 16:00 UTC. Integrations that use search in an agent loop should therefore check call volumes before that date and budget for the corresponding expense.
The model and search are separate parts of the release with different terms. The name deepseek-flash refers to model selection, while Search & Fetch billing applies to calls to search endpoints. When planning a budget, separate generation costs from search costs, and record how often each tool is called in agent tests. This will show what share of the expense goes toward retrieving web data.
The takeaway for teams: the update already affects users of the existing model name, and the future start date for search billing is known in advance. Check quality on your typical tasks, record behavior before and after the update, and recalculate scenarios in which an agent regularly calls Search & Fetch.