DeepSeek refreshes Flash and opens separately billed web tools

An automated research brief can change even when its configuration still says deepseek-flash. DeepSeek’s October 17, 2026 update moves that name to V4.1-Flash-1017 across the API, web and app. The calling method and model prices remain unchanged. The same announcement opens standalone Search and Fetch APIs, so a research worker now has two different kinds of usage to budget. Dated changelog.

The model name stays, so check the work it produces

DeepSeek reports stronger benchmark results, shorter outputs and lower task runtime. Those are vendor findings, not a promise that your next brief will be faster or more accurate. An unchanged API name also does not establish unchanged behaviour.

For an illustrative check, keep three briefs whose sources you have already reviewed: a price comparison, a launch summary and a question with conflicting evidence. Run the same instructions and inspect whether each answer retains the important qualification. Record the usable result, rather than rewarding a shorter answer that loses an exception. We have not run this comparison.

Count web requests separately from tokens

The published Flash rates per million tokens are $0.30 input on a cache miss and $1.20 output at peak times, or $0.15 and $0.60 off-peak. Cached inputs have separate rates. Weekends are off-peak; the pricing page defines weekday peak windows in UTC. Current model rates.

Search finds pages; Fetch retrieves one page as Markdown. Each costs $0.75 per 1,000 successful requests. Thus an illustrative 1,000 searches plus 2,000 page fetches costs $2.25, before model tokens or an applicable waiver. Empty results, failures and timeouts are not billed. Both endpoints have separate five-request-per-second account limits, and Fetch also has a per-domain limit. Tool guide and billing rules.

The current guide waives specified server-side tool fees through DeepSeek Harness, Claude Code or Codex until billing begins November 1 at 00:00 Beijing time. Do not apply that exception to every direct API request or to model tokens.

Start with a known brief and a small request budget. Our Perplexity guide helps choose the source workflow; the AI output evaluation guide helps decide whether the answer is ready to use.