- Anthropic introduces Claude 3, a powerful GenAI model family, showcasing enhanced analysis and forecasting abilities.
- Claude 3's unique multimodal capabilities allow it to analyze both text and images, outperforming several existing models including GPT-4.
- Despite its advanced features, Claude 3 avoids artwork generation and faces challenges with low-quality image processing and spatial reasoning tasks.
March 6, 2024: Anthropic, a leading AI startup backed by significant investment from Google and venture capital, has announced the release of its latest GenAI technology, Claude 3.
This new family of models, comprising Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus—the latter being the most potent—claims to outdo the performance of OpenAI’s GPT-4, especially in areas of analysis, forecasting, and specific benchmark tests against other major AI models.
Claude 3 marks Anthropic’s entry into multimodal GenAI, capable of understanding both text and images.
This advancement enables Claude 3 to process various document types, including photos, charts, graphs, and technical diagrams, distinguishing it from some versions of GPT-4 and Google’s Gemini 1.0 Ultra.
However, Anthropic has chosen to limit certain functionalities, such as artwork generation and identifying individuals in images, to navigate ethical and legal considerations.
The model is designed to provide more expressive and engaging responses, better follow multi-step instructions, and produce outputs in structured formats.
It also boasts an expanded context window, enabling a deeper understanding of inputs and generating more contextually rich responses.
Anthropic’s latest AI, Claude 3, has shown significant improvements in providing accurate responses, which is essential for businesses serving customers across various industries.
By evaluating Claude 3’s performance on a broad range of complex, factual questions designed to expose weaknesses in current AI models, Anthropic observed a notable increase in accuracy. Specifically, Claude 3’s Opus version has doubled the accuracy rate on these difficult questions compared to its predecessor, Claude 2.1, while also reducing the instances of incorrect answers.
Here’s the model pricing:
| Model | Claude 3 Opus | Claude 3 Sonnet | Claude 3 Haiku |
|---|---|---|---|
| Cost | [Input $/million tokens | Output $/million tokens] $15 | $75 |
| Context window | 200K* | 200K | 200K |
Yet, Claude 3 is not without its limitations. It cannot search the web, which restricts its answers to pre-August 2023 data, and shows less fluency in certain languages compared to English.
Anthropic acknowledges existing challenges like bias and hallucinations but promises regular updates to improve Claude 3’s capabilities.
Anthropic’s vision extends beyond the current iteration of Claude 3, aiming to pioneer an advanced AI self-teaching algorithm.
This ambition aligns with efforts to develop AI systems capable of performing a wide range of tasks, from administrative assistance to creative endeavors.
