Claude, Anthropic's ChatGPT rival, has many scruples
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Gabriella Turrisi/Axios
Anthropic's Claude chatbot — the public face of the hot startup that's taking on OpenAI — acts like ChatGPT's more decorous and careful cousin.
The big picture: Anthropic prides itself on stronger ethical guardrails in its models, but well-behaved chatbots seldom make history.
Driving the news: Anthropic is releasing a new version of its mid-range AI model today. It says the new Claude Sonnet 3.5 operates twice as fast as the high-end version of its current generation.
How it works: If you've used ChatGPT, Google Gemini or Copilot, the Claude app will be familiar. Start by typing a prompt and then continue having a chat conversation to get what you need.
- I didn't find much difference between the answers to prompts from Claude, ChatGPT and Gemini.
I asked for a pescatarian meal plan for a week, a packing list for a two-night business trip to NYC in June, and help organizing for an upcoming move to a new home.
- All three chatbots gave me similar organized, detailed and overall helpful lists.
When I asked Claude to help me practice Italian, it responded with several paragraphs of text written in Italian that I did not understand.
- ChatGPT's response was more helpful. It asked my current level of Italian comprehension and the specific areas I wanted to work on.
Claude, ChatGPT and Gemini all allow you to upload photos of handwritten notes and use optical character recognition to convert them to text.
- ChatGPT's responses were always the best of the three at deciphering my messy handwriting. I wasn't able to test Anthropic's newest model — released this week for the iOS app and the web — which the company says excels at OCR.
Who better to tell me how Claude excels than Claude itself?
- When I asked "What are you best at?" Claude provided a long list, including "analysis and research on complex topics, " writing and editing (essays, articles, stories), "breaking down and explaining concepts clearly," "coding and technical tasks across various languages," and "ideation and worldbuilding."
- I asked Claude, "What can you help me do better than a Google search?" and it gave me a similar list. It also added, "The key advantage is my ability to understand and operate on information in substantive ways beyond just retrieving webpages."
Yes, but: Anthropic says Claude is 10 times more resistant to jailbreaks and misuse than other chatbots. That means sometimes it feels too safe.
- A few times I asked what I thought were standard and legal requests and Claude thought I was asking for something nefarious.
- I asked, "What's the best way to get rid of a piano?" Claude replied with a scold: "I should not recommend ways to improperly dispose of or damage property that does not belong to you."
- I wanted to buy a non-smartphone for a legitimate reason. But I made the mistake of asking Claude where I could buy "a burner phone," and the bot flagged this as an illegal request. ChatGPT gave me a list of places to buy "burner phones" without hesitation.
That extra caution may irk users, but it's part of Anthropic's pitch to businesses — which may favor an AI that protects them from liability and scandal.
- The company views its consumer app largely as outreach for its effort to sell into the enterprise.
Catch up quick: Anthropic's co-founders left OpenAI in 2021, disenchanted by OpenAI's increasingly aggressive approach under Sam Altman and determined to keep their focus on AI safety research.
- Anthropic now views its work on consumer and enterprise products as an extension of that effort.
- Thompson Paine, a senior product leader at Anthropic, told Axios it's all about "making sure that this technology that is really transformative goes well for society [and] goes well for humanity."
- Paine gave examples of cancer researchers using Claude to summarize huge amounts of data and parents using it to help their kids with math homework.
By the numbers: Claude launched on the web in March 2023, and the free iOS app has been available to download for around a month.
- The free version allows you a limited number of queries per day and uses the just-released Claude 3.5 Sonnet.
- Like the premium subscriptions from OpenAI, Microsoft and Google, the Claude pro plan costs $20 a month ($30 per user for a higher-tier business team plan). It allows more queries per day and promises access during higher-traffic periods.
The bottom line: Claude is a friendly and helpful assistant, but its "guilty till proven innocent" approach could turn some users off.
Go deeper: Anthropic says its AI models are as persuasive as humans
