Web search on the Anthropic API
anthropic.com
anthropic.com
I've built a lot of LLM applications with web browsing in it. Allow/block lists are easy to implement with most web search APIs, but multi-hop gets really hairy (and expensive) to do well because it usually requires context from the URLs themselves.
The thing I'm still not seeing here that makes LLM web browsing particularly difficult is the mismatch between search result relevance vs LLM relevance. Getting a diverse list of links is great when searching Google because there is less context per query, but what I really need from an out-of-the-box LLM web browsing API is reranking based on the richer context provided by a message thread/prompt.
For example, writing an article about the side effects of Accutane should err on the side of pulling in research articles first for higher quality information and not blog posts.
It's possible to do this reranking decently well with LLMs (I do it in my "agents" that I've written), but I haven't seen this highlighted from anyone thus far, including in this announcement.
Would be interesting to see comparisons for custom web search RAG vs API. I'm assuming that many of the search "params" of the API could be controlled via prompting?
Interesting, I'm taking isotretinoin right now and I've found it's more interesting and useful to me to read "real" experiences (from reddit and blogs) than research papers.
I'm not trying to challenge your point, I am genuinely curious.
The API request notably includes the exact text it cites from its sources (https://docs.anthropic.com/en/docs/build-with-claude/tool-us...), which is nifty.
Cost-wise it's interesting. $10/1000 queries is much cheaper for heavy use than Google's Gemini (1500 free per day then $35/1000) when you'd expect Google to be the cheaper option. https://ai.google.dev/gemini-api/docs/grounding
> Web search is available on the Anthropic API for $10 per 1,000 searches, plus standard token costs for search-generated content.
> Each web search counts as one use, regardless of the number of results returned. If an error occurs during web search, the web search will not be billed.
https://docs.anthropic.com/en/docs/build-with-claude/tool-us...
> Web search results in the conversation are counted as input tokens on subsequent completion requests during the current turn or on subsequent conversation turns.
Yes, that's clear.
- If you show users text generated by Gemini using Google Search (grounded Gemini), you must display a provided widget with suggested search terms that links directly to Google Search results on google.com.
- You may not modify the text generated by grounded Gemini before displaying it to your users.
- You may not store grounded responses more than 30 days, except for user histories, which can retain responses for up to 6 months.
https://ai.google.dev/gemini-api/terms#grounding-with-google...
https://ai.google.dev/gemini-api/docs/grounding/search-sugge...
Popularity will grow even more, hence competition will increase and prices will change eventually
The Bing Search API is priced at $15/1k queries in the cheapest tier, Brave API is $9 at the non-toy tier, Google's pricing for a general search API is unknown but their Search grounding in Gemini costs $35/1k queries.
Search API prices have been going up, not down, over time. The opposite of LLMs, which have gotten 1000x cheaper over the last two years.
As I discovered recently, and much to my surprise, Google does not offer a "general search API", at least not officially.
There is a "custom search" API that sounds like web search, but isn't: it offers a subset of the index, which is not immediately apparent. Confusing and misleading labeling there.
Bing offers something a bit better, but I recently ended up trying the Kagi API, and it is the best thing I found so far. Expensive ($25/1000), but works well.
Try it and you'll see — there is no official Search API and the Custom Search API is quite poor and not usable in most scenarios.
Curious... I can understand that this may be a defensive action, however, feels unenforceable. And in some cases impractical for the user, after seeing this I may keep looking for alternatives for example because it's not clear to me if I have a chat history that has the search results in one of the messages, do I have to have a kind of mechanism to clean those out or something?
As far as I know, the pricing really hasnt gone down over the years. If anything it has gone up because Google is increasingly making it harder for these providers
For 100 results per query, serper.dev is $2/1000 queries and Bright Data is $1.5/1000 queries.
I always set 100 results per request from Bright Data, and I can see my bill indeed says `SERP Requests: x reqs @ 1.5 $/CPM` (where `x` is the number of requests I've made, not x * 100).
https://docs.brightdata.com/scraping-automation/serp-api/faq...
For serper.dev, they consider 10 results to be 1 "credit", and 20 to 100 results to be 2 "credits". They bill at $50/50,000 credits, so it becomes $1/1000 requests if you are okay with just 10 results per request, or $2/1000 requests if you want 100 results per request.
(Both providers here scale pricing with larger volumes, just trying to compare the easiest price point for those getting started.)
Gemini is $30/1000
So Anthropic is actually the cheapest.
For context, exa is $5 / 1000.
The rate limits of the upstream engines are fine for personal use, and the benefit is it uses the same browser you do, so results are customized to your search habits out-of-the-box (or you could use a blank browser profile).
Presumably this is using Brave under the hood, same as Claude's search feature via the Anthropic apps?
Google Gemini has some: https://ai.google.dev/gemini-api/docs/grounding/search-sugge...
OpenAI has some rules too: https://platform.openai.com/docs/guides/tools-web-search#out...
> "When displaying web results or information contained in web results to end users, inline citations must be made clearly visible and clickable in your user interface."
I'm used to search APIs coming with BIG sets of rules on how you can use the results. I'd be surprised but happy if Anthropic didn't have any.
The Brave Search API is a great example of this: https://brave.com/search/api/
They have a special, much more expensive tier called "Data w/ storage rights" which is $45 CPM, compared to $5 CPM for the tier that doesn't include those storage rights.
When your intelligent email client uses Gemini to identify the sender of an email as someone in the industry your B2B company serves, deciding to flag the email as important, where is that HTML supposed to go? Where does it go in a product that generates slide show lesson plans? What if I'm using it to generate audio or video? What if a digital assistant uses Gemini as a tool a few dozen times early in a complex 10,000 step workflow that was kicked off by me asking it to create three proposals for family vacations complete with a three 5-minute video presentations on each option? What if my product is helping candidates write tailored cover letters?
It's bad optics for a company just ruled to have acted illegally to maintain a monopoly in "general search services and general text advertising," but worse, it lacks imagination.
https://docs.anthropic.com/en/docs/build-with-claude/tool-us...
Also I'm quite sure that they don't use vector embeddings for web search, its purely on text space. I think the same holds for all LLM web search tools. They all seem to work well -- maybe we don't need embeddings for RAG and grepping works well enough?
Does this mean there are certain sites that the search tool simply can't access?
Even if you search for videos on a topic with hundreds of highly relevant YouTube videos, and explicitly ask it to find YouTube videos, it still says it found other relevant stuff, but not YouTube videos.
Maybe Google search revenue moves from ads to more towards B2B deals for search API use.
Clearly web search ranking is hard after decades of content spam that's been SEO optimized (and we get to look forward to increasing AI spam dominating the web in the future). The best LLM provider in the future could be the one with just the best web search ranking, just like what allowed Google to initially win in search.
The LLMs can access the web, but they can't yet understand it in a structured, evaluative way.
What’s missing is a layer of engineered relevance modeling, capable of filtering not just based on keywords or citations, but on deeper truth alignment and human utility.
And yes, as you mentioned, we may even see the rise of LLM-targeted SEO—content optimized not for human readers, but to game LLM attention and summarization heuristics. That's a whole new arms race.
The next leap won’t be about just accessing more data, but about curating and interpreting it meaningfully.
Why do you think it is limited? Imagine you show a link with details to an LLM and ask it if it is trustworthy or high quality w.r.t the query, why can't it answer it?