DeepSeek updates V4.1 Flash and opens a paid web search API for agents

What changed

According to DeepSeek's API change log, the deepseek-flash model name now points to the new version, and the web and app versions were updated too. The API calling method and pricing are unchanged.

DeepSeek reports higher scores on coding, agent and vision tests. One example: Terminal-Bench 4.0 rose from 31.2 to 42.3. It also says the model finishes tasks faster and writes shorter outputs, which lowers cost. These are the company's own results, not independent tests. The model card lists 763 billion parameters, and the weights are released under the MIT licence.

The new Search and Fetch APIs

The Search and Fetch guide describes two tools for agents and retrieval systems. Search returns structured web results from DeepSeek's own search engine. Fetch returns a single web page as Markdown text.

Each costs $0.75 per 1,000 successful requests. Failed requests, empty results and timeouts are not billed. Each account is limited to 5 requests per second per API. When the DeepSeek API is used through DeepSeek Harness, Claude Code or Codex, these server-side calls are free for now; DeepSeek says billing starts at 00:00 Beijing time on 1 November 2026.

What it costs in practice

On the pricing page, Flash costs $0.30 per million input tokens without a cache hit and $1.20 per million output tokens at peak rates. Off-peak rates are half. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; the second window overlaps the UK and European working morning.

A hypothetical example, using our own assumptions: an agent runs 2,000 research tasks a month, each with one search, one page fetch, 20,000 input tokens and 2,000 output tokens. The tool calls cost 4,000 × $0.75 / 1,000 = $3.00. The tokens cost $12.00 for input and $4.80 for output at peak rates. That is about $19.80 a month at peak, or $11.40 if all the token use is off-peak. Real costs depend on prompt size, caching and retries.

Data questions for UK and EU businesses

Price is only part of the decision. DeepSeek's developer terms say the platform is operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd. and governed by the law of mainland China. DeepSeek's privacy policy, which covers its own apps, websites and related services, says it collects, processes and stores personal data in the People's Republic of China; under the platform terms, the personal data a developer provides as a data subject is handled under that policy. That policy states it does not cover the personal data of end users of apps built on the platform: for those, the developer is the controller. Under the terms, if your app serves the public, you must disclose your own personal-information processing rules to end users and have their consent or another legal basis for collecting the data and for delegating its processing to DeepSeek.

Italy's regulator has acted before. In January 2025 the country's data protection authority, the Garante, ordered an urgent limitation of the processing of Italian users' data by the DeepSeek chatbot service. That order concerned the consumer chatbot, not this API update.

The MIT licence gives another option: run the model on infrastructure you choose, so prompts do not go to DeepSeek's API. At 763 billion parameters, that means server-class hardware or a hosting provider, not a laptop.