Claude 3¶
Anthropic, an AI startup backed by Amazon and Google, has launched the brand-new Claude 3 large language model, which includes three versions: Haiku, Sonnet, and Opus. According to official data, Opus outperforms both GPT-4 and Gemini.
Claude 3 Models¶
-
Claude 3 Opus
Opus means "work" in Latin. Anthropic describes it as their most intelligent product, with an understanding level close to that of humans, and it currently performs the best in several benchmark tests. In ten commonly used model capability assessments listed by Anthropic, Opus scored higher than GPT-4, even in visual recognition abilities.
-
Claude 3 Sonnet
Sonnet refers to a "fourteen-line poem" in English. It strikes an ideal balance between intelligence and speed, making it particularly suitable for high-traffic tasks. For most tasks, Sonnet is twice as fast as Claude 2 and Claude 2.1, while also demonstrating a higher level of intelligence.
-
Claude 3 Haiku
Haiku is a transliteration of the Japanese word for "haiku." It is the most cost-effective model among the three, providing nearly instantaneous responses.
Comparison of Claude 3 and GPT-4¶
Claude 3 now supports multimodal capabilities similar to GPT-4. Opus scored 59.4% on the MMMU, surpassing GPT-4V and matching Gemini 1.0 Ultra. If you want to upload photos or images for AI analysis, data shows that Claude 3's capabilities are better than those of GPT-4.
The Claude series models also boast an advantage with their long context windows. The updated full series of LLMs offers a context window of 200K and supports input of over 1 million tokens. How long is 1 million tokens? It's approximately the length of the English version of "Harry Potter and the Deathly Hallows."
However, in terms of pricing, the most capable Claude 3 is significantly more expensive than GPT-4 Turbo. GPT-4 Turbo charges $10/30 per million tokens for input/output, while Claude 3 Opus charges $15/75, making its output 2.5 times more expensive.