Anthropic AI overtakes itself: Better image recognition, better coding
Anthropic has launched the first version of its AI assistant Claude in version 3.5 Sonnet. It is supposed to be faster and smarter for the same price.
(Image: Anthropic)
Anthropic's AI assistant Claude is now available in an improved version. Claude 3.5 Sonnet replaces Claude 3 Sonnet. Sonnet is the "middle" version of Claude: Its smaller sibling Haiku is cheaper but not as "intelligent", while Opus is more expensive to use but more "intelligent" – basically. Because now Claude 3.5 Sonnet is supposed to be cheaper and smarter than Claude 3 Opus.
Faster too: the new Sonnet should solve tasks in half the time that the old Opus takes. This is particularly attractive for chatbot operators who want to outsource customer support from knowledgeable humans to calculating AI. The provider plans to release Claude 3.5 in the course of the year in the cheaper Haiku version and the more powerful Opus version. Claude has also been available in Europe for around five weeks.
The provider is particularly proud of the improvements achieved in internal tests when programming software improvements, i.e. correcting errors or adding new functions to existing open-source code. Where Claude 3 Opus masters 38 percent of the tasks set, Claude 3.5 Sonnet succeeds in 64 percent of cases.
Anthropic also claims to have significantly improved image analysis, especially when interpreting tables and graphics. The new AI assistant is also better at transcribing text in imperfectly captured images. In a series of other benchmark tests, Claude 3.5 Sonnet outperforms its competitors GPT-4o and Gemini 1.5 by several percentage points. These tests are standardized and not very meaningful for the actual handling of chatbots.
Same price
Claude 3.5 Sonnet is available in a free version for internet browsers and as an iOS app, albeit with rate limits. For paid use, these limits are set five times higher. As with the previous version, this costs three US dollars per million input tokens and 15 dollars per million output tokens, with a context window of 200,000 tokens. This describes how much Claude can process at once.
"Our goal is to significantly improve the trade-offs between comprehension, speed and cost every few months," says Anthropic, which counts Amazon.com and Google among its investors. At the same time, Anthropic is working on additional offerings to make AI more usable.
One new feature is a workspace called Artifact. It displays the content that Claude has just created at the user's request in a separate window. The content can then be edited and reused. Currently only by the respective user, but in future also by colleagues. Claude could lose its memory exemption in the future: The developers are researching a feature called Memory, with which Claude remembers previous interactions with a specific user and their preferences; this should increase personalization and efficiency.
(ds)
