Anthropic Unveils New Model “Claude 3”, More Powerful than GPT-4
Anthropic, backed by investments from Amazon and Google, has announced the release of its Claude 3 series models, claiming superiority over all competitors, including GPT-4. These models are touted as the fastest and most powerful, even demonstrating “human-like” abilities in certain tasks.
“We are pleased to announce the launch of the Claude 3 series models today, which sets a new industry standard for a wide range of cognitive tasks,” stated Anthropic on its official website.
The Claude 3 series models released by Anthropic include three models: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. These models are arranged in ascending order of performance to provide users with options based on their specific needs and costs. Currently, Opus and Sonnet can be used on Claude.ai and the Claude API, with Opus requiring a monthly subscription of $20 for Claude Pro, while Haiku is set to be released soon.
Anthropic has provided different positioning for each model:
Claude 3 Opus:
This is the most advanced model from Anthropic, possessing abilities and fluency that are close to human understanding. It is designed for extremely complex tasks and open-ended prompts.
Claude 3 Sonnet:
This model strikes a balance between intelligence and speed, offering better performance and cost-effectiveness compared to similar products. It is designed for high durability required in large-scale AI usage.
Claude 3 Haiku:
This is the smallest and fastest model, with near-instantaneous response capabilities. It can quickly answer simple questions and is used for real-time interactions with users.
Anthropic claims that in most tests, Opus, the most powerful model in this release, outperforms major AI models in the market, including university-level expert knowledge (MMLU), graduate-level professional reasoning (GPQA), and basic mathematics (GSM8K), surpassing GPT and exhibiting understanding and fluency close to human levels in complex tasks.
In terms of visual capabilities, the Claude 3 series models also perform on par with their competitors, handling complex visual content such as photos, charts, and technical diagrams.
Anthropic points out that more than half of the knowledge repositories of many customers comprise various types of visual content, such as PDFs, flowcharts, and slides. They are pleased to offer this new modality to their customers. It is worth noting that while the Claude 3 series models can process images, they do not generate image content.
According to CNBC, Anthropic states that Claude 3 can handle approximately 150,000 words (200,000 tokens) of text, equivalent to books like “Moby-Dick” or “Harry Potter and the Deathly Hallows.” Previous versions could only process about 75,000 words.
In the data revealed by Anthropic, the lightweight Haiku model can process research papers with complex charts and approximately 10,000 tokens in less than 3 seconds.
Regarding pricing, Anthropic offers Opus at an input cost of $15 per million tokens and an output cost of $75 per million tokens. This price is significantly higher than GPT-4 Turbo, which charges $10 per million tokens for input and $30 per million tokens for output, perhaps indicating Anthropic’s confidence in their own model.
Reducing the Model’s “Hallucination Rate”, Anthropic Strives to Make Claude 3 Safer
Chatbots are susceptible to being misled or providing fabricated answers due to inadequate understanding of questions. This phenomenon, known as “hallucination,” is considered a possible source of misinformation.
Anthropic attempts to address this issue with Claude 3, claiming that in tests involving a series of challenging and complex questions, the Opus model’s accuracy has more than doubled compared to previous models, significantly reducing the proportion of incorrect responses. However, they admit that completely eliminating this problem is not easy. “Achieving a hallucination rate of zero is extremely difficult,” said Daniela Amodei, President of Anthropic.
“None of the models are perfect, and I think it’s important to make that clear,” emphasized Amodei. “We are doing our best to make the models safer and more powerful, but there are still times when fabricated responses occur.”
Source: Anthropic, Bloomberg, CNBC
Edited by: Mei-Hsin Lin