What changed
DeepSeek announced V4.1 Flash 1017 on October 17. Its changelog says the website and app were updated, and the existing deepseek-flash API name now selects this version. The API calling method and model pricing remain unchanged, according to DeepSeek.
The company reports improvements in coding, reasoning, agent tasks, and visual understanding, with shorter outputs and faster task completion. Those are provider-reported results. We have not run the model or independently checked whether its gains carry over to a particular project.
Search and Fetch have their own bill
Search finds web results; Fetch retrieves a page. Each endpoint costs $0.75 per 1,000 successful requests that return results or page content. Failed requests, empty search results, and timeouts are not billed. Requesting more search results does not change the per-request price.
DeepSeek currently waives these fees for server-side Search and Fetch tool calls made through its API in DeepSeek Harness, Claude Code, and Codex. Billing for those calls begins October 31, 2026, at 16:00 UTC. That waiver does not cover model inference charges.
Open weights still need deployment work
DeepSeek publishes the model weights under MIT terms. The official repository lists about 510 GB of files. That storage figure does not establish runtime memory needs or laptop suitability. Check the actual format, serving software, and hardware requirements before planning a local deployment.
Our take
Editorial analysis: compare the old and new versions on the same useful task, with the same tools, permissions, and success criteria. Keep the outputs and inspect the finished work. A faster answer is helpful only if the result meets the requirements.
Count model usage and web-tool calls together when estimating a workflow's cost. Check what queries leave your machine, limit access to private files, and read the original sources behind a generated summary. Retrieved pages can supply evidence; their instructions should not automatically control your agent.
Source timing
This is October 17 catch-up coverage, source-checked October 18. It is a separate update from the September 10 release discussed in our earlier DeepSeek briefing.