The problem with working in a hyper-competitive industry is that there’s maximum motivation to roll out new features, even when those features aren’t ready yet or adequately tested. OpenAI received a lot of praise for its Deep Research function through its $ 200-a-month model, but my experience with it found it to hallucinate heavily — i.e., make a lot of stuff up. And the ability to write a Ph.D.-level research paper becomes a lot less valuable if you can’t trust what it’s saying to be real.

A similar issue has existed with AI search tools, which use large-language models to search the web and bring back links for you. In theory, this is a fantastic function, but, as the Columbia Journalism Review recently noted, these tools don’t cite well, speculate frequently, invent links, and often summarise badly. OpenAI, Perplexity, and XAI know this — this article wasn’t news to them. But, rather than wait and make the product work well, they are incentivized to release quickly and early and then fix issue later.

Anthropic has been the one major outlier here. Their model, Claude, has not simply been among the most powerful and useful on the market but also built more conservatively, with Anthropic ensuring that it works well and can’t be easily jailbroken before releasing its features. For months, Claude has had a developer preview version of a web search function, but they put no priority on bringing it to market. But this week, it finally went live — at least for American users — and the results seem pretty impressive.

The most interesting point is that this isn’t a separate interface or system from regular Claude. Instead, when enabled in the settings menu, Claude can search the web but only does so when addressing more complex queries than it thinks it can answer or when searching for more up-to-date information. When I asked Claude about improving the scrolling speed on my camera, it made some helpful suggestions for improving it through an upgraded SD card, doing so from its existing training data without taking the time to search the internet. However, when I provided it with a link to the SD card I’m using, it could open that link and give me information specifically based on it in a way it couldn’t before.

Similarly, when I asked it whether there was a Windows app for transferring pictures from my camera to my computer over WiFi, it replied quickly without searching the web, saying it didn’t know for sure, and asked if I would like it to search the web to see if there was more updated information. When it did, it provided a list of links and citations; and, unlike some other tools, it defaulted to being concise, relevant, and giving me a quick range of options. It also hasn’t hallucinated any links.

This is a unique way of handling web searches for an LLM — which, it should be noted, seemed to be powered by Brave Search’s API — and though some won’t like it, it should help both to increase the speed of average responses and minimize hallucination. At the top of the result, Claude lists all the pages it searched but doesn’t clutter the screen like Perplexity, but instead keeps them in a fold-down menu. The one problem is that it can require more encouragement than necessary.

When I told Claude that I was “looking to buy a slim, stylish, premium 13″ laptop,” it produced a list of options based on its training data — which ran until October 2024 — and it was only when I explicitly asked it to search for all options for me that it went online. By comparison, the leading AI search engine, Perplexity — which can use Claude as it’s AI model — just instantly started searching for options. I preferred the search straight out of Claude — it summarised the information better and was more concise — but having to nudge it to do what you want is frustrating.

Claude’s Web search is now available for all paid Claude users through the feature preview function, which will be available to free and international users in the near future.