#

DeepNews

The DeepNews endpoint is our state-of-the-art AskNews agent, capable of self-reflection, on-the-fly learning, and deep tangential iterative research into the AskNews news knowledge graph, Google, Wikipedia, X (Twitter), Reddit, send emails, and more. This agent is good for tasks that require complex analysis and high-level strategic planning. For example, monitoring geopolitical risk, investigative journalism, making event forecasts for prediction markets, fact-checking, and much more.

You can use most of the latest models with DeepNews, including Anthropic models, Google models, OpenAI models, and open source models. The currently recommended model parameter is claude-sonnet-5 or claude-opus-4-6. For a fast model, we recommend open-source-best (this model changes accordingly with our evaluation of the latest open source models). See the models list below and the model type. Each type has a different cost (see rate limiting costs for details). The full list of available models for DeepNews is also available at the updated API reference. If you want to request model support for a specific model, please reach out to us at contact@asknews.app.

Loading models...

This endpoint is also available as a tool on the MCP server.

Wanna test it out or request a new model? You can come to our Discord server to talk directly to the bot, or ask us to integrate new models

Detailed and updated response structures are always available in the API reference

#

Querying the DeepNews endpoint

The DeepNews endpoint can be used with the typical OpenAI SDK by setting base_url to https://api.asknews.app. However, if you want fully structured source objects in addition to your response, use one of the AskNews SDKs as shown here:

#

AskNews SDK

For engine="v2.0" streaming, DeepNews emits typed content-block events for text, thinking, and tool input, plus separate source messages when return_sources=True. Keep side-effecting tools such as email out of sources unless you explicitly want the agent to use them.

#

Controlling the output

The following parameters allow you to make decisions about how deep the agent is allowed to search, and how the response is returned to you:

• engine The DeepNews engine version to use. Set to "v2.0" for the typed streaming format shown above. It uses the same next-generation agent as "v1.5", but streams thinking, tool-use, text, and JSON deltas as typed content-block events. The legacy engine is "v1". Defaults to "v1.5".
• sources The sources to search in. You can choose from asknews, google, wiki, x, reddit, charts, graph, email. By adding these, DeepNews will intelligently decide which tool is best for various parts of each research project. The ordering of these sources is the prioritization that you want DeepNews to consider when choosing tools. The email source allows DeepNews to send an email as part of its research workflow (e.g. sending a summary or report to a specified recipient).
• max_parallel_tool_calls The maximum number of tools the agent is allowed to call in parallel at each research step (up to 4). Higher values can significantly speed up research by allowing the agent to search multiple sources simultaneously, while lower values force the agent to reflect on each tool's response, improving accuracy of subsequent searches.
• search_depth The number of levels deep you want to force the agent to search. Beyond which, the agent can choose if it wants to continue searching or if it is done (up to max_depth)
• max_depth The maximum number of levels deep the agent is allowed to search.
• return_sources Whether to return structured source objects with the response. For stream=True, source objects are yielded as CreateDeepNewsResponseStreamSource items.
• only_cited_sources Whether to return only sources that are cited or referenced in the generated response content.
• enable_source_pruning Whether to prune sources that are deemed irrelevant to the final response. Leave this False when you want to inspect all returned evidence.
• stream Whether to stream the response or not. If stream=True, the response is a generator that yields each typed stream chunk as it is received. If stream=False, the response contains the completed message.
• model The model to use for the response. See the list at the top of this page for the available models. If you want another model, please reach out with your request at contact@asknews.app.
• filter_params (Optional) The AskNews filter parameters (which can be manually controlled or automatically populated as shown in the AutoFilter documentation). This allows you to filter the sources that are returned in the response based on various criteria such as date, source type, country, sentiment, etc.

Detailed and updated response structures are always available in the API reference