Vercel added an accelerated DeepSeek Flash route to AI Gateway on 7 November. Search and page retrieval tools arrived the day before, when Vercel also made Tracing generally available. For teams building research agents, the releases give them separate points to measure generation, check the evidence their agents retrieve, and investigate slow requests.
Coverage started with the preceding 24 hours and expanded to a 72-hour lookback for the related 6 November releases. The announcements show dates without exact publication times.
An accelerated route for the October DeepSeek Flash update
Vercel added DeepSeek V4.1 Flash 1017 Ultrafast through Fireworks. It serves the 17 October model update through an accelerated endpoint. Vercel reports that it can exceed 1,000 generated tokens per second. That figure describes generation throughput, and the announcement gives no comparable end-to-end timing for a workflow that makes several tool calls.
Benchmark the route on representative prompts. Record time to first token, generation time and total task time separately so you can see where faster inference actually helps.
https://vercel.com/changelog/deepseek-v4-1-flash-1017-ultrafast-now-available-on-ai-gateway
Search and page retrieval are separate tool calls
DeepSeek Search returns titles, URLs and summaries. DeepSeek Fetch retrieves page content as Markdown. Vercel says both tools work with models that support tool calling through AI Gateway. The AI SDK helpers require version 6.0.321 or later on v6, or 7.0.160 or later on v7.
Use search to find candidate pages and fetch the pages that support the answer. Keep the source URL with each claim and treat fetched text as untrusted input.
https://vercel.com/changelog/ai-gateway-adds-deepseek-search-and-fetch-tools
Tracing is available on all Vercel plans
Vercel made Tracing generally available on 6 November, with KQL search in its dashboard and CLI. Retention is three days on Hobby and 14 days on Pro and Enterprise. New Vercel projects on Next.js 16.5 or later with a Node.js function get tracing after their first successful build.
Check sampling and retention for your own project. If your application records tool-call spans, compare them with model and application spans before attributing a slow run to inference.
https://vercel.com/changelog/vercel-tracing-is-now-generally-available