News from the IT community on October 17th: This morning, there were multiple consecutive changes at the API level in DeepSeek.

First of all, DeepSeek officially announced that Search / Fetch API has been officially launched. Both interfaces use DeepSeek's self-developed search engine to provide real-time web content for Agent and RAG applications.

The Search interface searches web pages according to the query criteria and returns structured results, while the Fetch interface grabs a single web page and returns the title and body of the content in Markdown format. Both use the original JSON API data, and Base URL is equivalent to api.deepseek.com. They both employ the same Bearer authentication method as DeepSeek API.

According to the DeepSeek support documentation, both Search and Fetch are billed based on the number of successful requests, at a rate of 5 yuan per thousand requests. Fees are only incurred when Search returns a non-empty result list, Fetch returns the page content, and the HTTP status code is 200. Failed requests, empty results, and timeouts are not charged. The number of result items returned by the requests does not affect the pricing.

Each account is limited to 5 queries per second for Search and Fetch; these limits are counted independently. There is also an additional limit based on domain name for Fetch. Exceeding these limits will result in a HTTP 429 error. The query for Search requests is limited to 128 characters, and the number of returned results can be selected between 1 and 50, with a default of 10. The results include a title, URL, a longer model-oriented abstract, an optional full text, the publication date, a reordering score, and an authority rating.

It is worth mentioning that within the first two weeks after the launch of Search API (before November 1, 2026), searches triggered through the Web Search tool of Claude Code or the web_search function of Codex are not subject to any Search fees.

At the model level, the robot account AI Tracker Bot on the X platform captured an entry labeled “deepseek-v4.1-flash-1017” in the official API list of DeepSeek, but it was quickly removed.

The account pointed out that the suffix “1017” is consistent with the naming convention for date snapshots prior to “DeepSeek”, such as “V4-Flash-0731” and “V4-Pro-0813”, therefore it may be pointing to new snapshots “V4.1” and “Flash”.

The account also updated that the deepseek-flash publicly available on DeepSeek API is still V4.1 Flash, and not a "1017" snapshot; the degree of consistency between its output and models with the same openness weight hosted by third parties is comparable to the consistency of V4 Pro between the two hosting parties. As of the time of publication, the official DeepSeek has not yet made any statements regarding the existence or changes to the "1017" snapshot.

As a background, DeepSeek officially released V4.1 and Flash on September 10th. These are the smallest models in its new model architecture series, which natively support multi-modal visual understanding. The model names have been changed to deepseek-flash for easy invocation. The previous generations, V4 Flash and V4 Flash Vision Exp, have been taken offline and temporarily routed to V4.1 and Flash.

Finally, DeepSeek also released the latest v0.2.1-rc1 version of DeepSeek Harness today, which is also the first candidate version of the 0.2.1 series. This version summarizes the main changes related to users and developers since v0.2.0-rc.2. The new features added in IT are as follows:

• Allows users to discover and install plugins from the "Custom Plugin Index Source".
• A new option has been added to the plugin management page: "Allow Agent to create plugins," which allows you to preserve drafts and enter creation mode; execution will only begin after a request is sent.
• Supports enabling, disabling, uninstalling plugins, and synchronizing dependencies after updates; when hot reloading is enabled, updated or reinstalled replaceable plugins take effect immediately. If hot reloading is not possible, a restart is prompted.
• Optional plugins support minimalist mode. The plugin list supports installing the Claude Code and Codex combination packages corresponding to the current release version as needed, and can also be used with headless, SDK, ACP, and profile.
• Added an experimental plugin's Session status recording interface, for plugins to write and read their own session records.
• New experimental Claude Code Mods compatibility layer has been added. The main goal at this stage is to verify that the functions of Claude Code Mods API are roughly a subset of those of the DeepSeek Harness plugin, rather than to provide users with actual full compatibility.
• New experimental developer tool package added, providing bidirectional navigation through session raw logs, chat groups, and text content, as well as an embedded NodeJS debugging tool; the embedded DevTools currently uses an English interface.
• New experimental thought process machine translation plugin available; it can be enabled on the plugin page by selecting Bing, Google, or the paid option DeepSeek Flash.
• New experimental Git Worktrees plugin has been added, which can be enabled on the plugin page. It is created by Agent and allows access to an independent Git checkout.
• The custom model API supports configuring the inference level of each model through UI, eliminating the need to manually modify the configuration files.
• pi-ai The routing supports updating system prompt words in conversations based on model capabilities, as well as the dynamic addition and removal of tools.
• Minimal mode supports reading images and automatically compressing long conversations. Web On the desktop version, you can manually compress the context using the / compact command.
• Global commands support reading from the AGENTS.md in the agents directory, which can be specified through the DSH_AGENTS_HOME directory.
• Added a new `working_directory` tool, and provided read and switch interfaces for the session working directory for TypeScript / Python SDK.
• SSH Added an independently executable helper. During runtime, it supports remote file operations, processes, terminals, sandboxes, and Node PTC.
• Web version supports logging in with a DeepSeek account when accessing from the local HTTP.
• Supports identifying keys and tokens in sent messages, and by default, they are masked for viewing as needed.
• New sessions support pre-filling of unsent prompts; files, directories, and session references are retained when restoring drafts or switching workspaces.
• Added a new option for setting the timing of closing work steps: you can choose to close them after answering is completed, or when the next regular message arrives.
• Markdown The file preview displays the YAML header metadata as a clear list of fields.

Related Reading:

• DeepSeek service failure for nearly 5 hours, official compensation for all API users' consumption within 7 days
• DeepSeek Website / API Service Restored to Normal, After Nearly 5 Hours of Interruption
• DeepSeek Website / API Service experienced partial interruptions today, with a message indicating "Server temporarily unavailable".