Last week, Anthropic introduced web search on the Anthropic API, giving developers access to Claude models that can tap into current information. Web search is available through the API for Claude 3.7 Sonnet, the upgraded Claude 3.5 Sonnet, and Claude 3.5 Haiku. With this announcement, Anthropic joins Google, Perplexity, OpenAI, Exa, and Tavily in offering AI web search through APIs. It comes at a time when Apple is considering a move to AI search on Safari, in response to Google’s antitrust loss in search.

Anthropic – founded by OpenAI’s ex-VP of Research Dario Amodei to work on AI built for safety – is relatively late to the game with AI search. Real-time access was not a built-in feature for Anthropic’s applications up until recently. In Nov 2024, Anthropic began allowing AI agents to connect to external tools and data sources – including live data streams and databases – through its open-source Model Context Protocol (MCP). (MCP has since been adopted by Anthropic rivals OpenAI and Google, as well as Microsoft, Cloudflare, Confluent, and Cursor, among others.) In Mar 2025, Anthropic enabled web search for paid subscribers through its interfaces. Now, with web search on the Anthropic API, developers can build Claude-powered applications/agents that tap into real-time data, while maintaining control through domain allow lists, domain block lists, and organization-level management.

Anthropic’s entry will help drive down the price of AI search. Its web search is currently priced at $10 per 1,000 queries (i.e. 1 cent per search) plus standard token costs. In contrast, Google’s web search is priced at $35 per 1,000 queries plus normal costs for Gemini, and OpenAI’s web search ranges from $25-50 per 1,000 queries. Other players are priced even lower, with Perplexity at $5-14 per 1,000 queries, Exa at $2.50-25 per 1,000 queries, and Tavily at $5-16 per 1,000 queries.

Notably, the AI players without roots in search are using partners to power their web search capabilities. Anthropic’s web search is reportedly powered by Brave, for instance, while OpenAI’s web search is reportedly powered by Bing and other undisclosed partners. Google obviously has its own search infrastructure to lean on. Perplexity – which was built to be an “answers engine” – maintains its own webpage index and rapidly orchestrates a combination of traditional web search and 3rd-party AI models to answer user questions. (Perplexity doesn’t train its own foundation models, although it does customize open-source models for its search and other needs.)

For Safari, Apple – which doesn’t yet have its own AI search – has been in partnership discussions with Perplexity, and will likely include OpenAI and Anthropic on its list as well. (For Siri, Apple partners with OpenAI’s ChatGPT and plans to add Google’s Gemini later this year; it also reportedly considered Anthropic, Perplexity, DeepSeek, and xAI’s Grok for Siri as well.)

We seem to be converging to a world where chatbot interfaces like ChatGPT are the places where people go for general searches – akin to how they used to go to Google but across an expanded set of use cases. Anecdotally, AI search is outperforming traditional search on many (though not all) use cases, including shopping, people, recipes, how-tos, movie recommendations, places to stream movies, and more.

On the surface, AI web search appears to represent a threat to Google. Apple recently revealed that Google searches on Safari were down in Apr 2025 for the first time in 22 years, which Apple believes to be the result of people turning to AI with their questions. With APIs for web search, it’s becoming increasingly possible for companies to offer high-quality search – both general search and within targeted domains – without Google.

On the other hand, Google has one of the best foundation model families out there – arguably the best, given its Gemini-2.5-Pro-Preview-05-06 is currently topping the leaderboards. It also has its search prowess and a still-dominant position in query traffic, with a 90% global market share. With AI Overviews and its improved search-results quality (an issue Google has struggled with), Google reports its overall query growth is actually up, including queries from Apple devices and platforms. For Google, better results can mean more searches because users “suddenly can ask questions that were too hard before.”

Google is facing the classic “innovator’s dilemma” problem, where it has to decide how and when to make the transition from traditional search to the less lucrative economics of AI web search. Despite Gemini’s models being “some of the most efficient models out there, including comparing to DeepSeek’s V3 and R1,“ the cost per AI-powered search is still probably a multiple of the cost of a traditional web search (it was 10x in 2023 but has come down since).

In addition to chatbot interfaces becoming widely used for general search, AI-powered search will also be broadly distributed across 3rd-party applications and websites. This is the race we are seeing here with web search via APIs. The use cases for integrating AI search are endless, from shopping, travel, and investment analysis to medical research, legal research and documentation reference.

From the perspective of the developers using the APIs, they are seeing an emerging set of increasingly useful and possibly less expensive tools. From the perspectives of the players in the race, this is the latest gold rush. As we have said before, the strategic flywheel of the AI era will be AI-powered personalization and the pot of gold at the rainbow’s end is the next big advertising business.

One of the key questions for AI players is what data can they use and how can they use it. Google, OpenAI, Anthropic, and Perplexity generally don’t use API data to train their models unless customers specifically opt in. However, Google’s “Grounding with Google Search” will store prompts and contextual information from customers, and the generated output, for 30 days “for the purposes of creating grounded results and search suggestions” and “for debugging and testing of systems that support grounding with Google Search.” Apple hinted at this issue in testimony at Google’s antitrust trial, saying that Google’s term sheet “had a lot of things Apple wouldn’t agree to and didn’t agree to with OpenAI” – although it’s unknown whether the terms in dispute were related to data privacy, exclusivity, co-branding, or something else.

We can imagine a scenario where AI players eventually negotiate custom deals with business users that allow them to tap search data to improve conversions and the end-customer experience. With enough of those deals under their belt, an AI player could track individuals across the internet to enable better personalization, more targeted ads, and tighter attribution between ad and purchase. In the meantime, for an AI player, offering AI web search via API can mean a lucrative revenue stream, greater developer familiarity and potential lock-in to their stack, and access to at least the metadata associated with the queries.

The coming changes to web search promise to be highly disruptive. AI web search breaks a lot of the business models on which the web has been architected, particularly when paired with AI agents. The power of Google’s traditional search franchise shaped the internet as we know it, as companies and individuals optimized for search discovery and ad placement (e.g. how webpages are structured, how content is written for SEO and searchability, backend infrastructure so fast-loading pages can rank higher).

Analysts are projecting a 25% drop in website traffic by 2026 – with websites in one study already seeing a loss of 18-64% from Google’s AI Overviews – due to the reduced emphasis on web links. Another study found that 60% of searches on traditional search engines are now “zero click” or ending without the user moving onto another destination. There may be fewer winners selected for the top summary in an AI search world, and smaller and newer websites may have virtually no chance at all. As AI web search becomes widespread, we may find ourselves with a very different-looking internet.

May 22 2025 (3 Shifts): OpenAI and AI players inch towards ads

Jun 21 2024 (3 Shifts): Reddit, ambient search, and the threats to Google