AI Chatbot Claude 2.1 Released: Supports Ultra-Large Inputs of 200K Tokens, Outperforming GPT-4 Turbo
-
On November 23, Anthropic, a competitor of OpenAI, launched the Claude 2.1 chatbot and its corresponding AI model. Claude 2.1 supports inputs of up to 200,000 tokens, improves accuracy, and introduces a new system prompt feature.
According to the official introduction, Claude 2.0 previously supported only 100,000 tokens per conversation, while Claude 2.1 now supports up to 200,000 tokens, equivalent to 150,000 words or a 500-page document. This surpasses OpenAI's GPT-4 Turbo, which supports 128,000 tokens (Note: On average, 1 token corresponds to 4 English characters, and 100 tokens roughly equal 75 English words).
Anthropic stated that since the launch of Claude's chat service this year, millions of users have utilized it for translating academic reports, drafting business plans, or analyzing complex contracts. Many users requested longer token limits when handling lengthy documents, prompting Anthropic to double the input capacity.
The company claims that with Claude 2.1, users can input content equivalent to "an entire book" for more detailed summarization, Q&A, trend analysis, and comparisons.
In terms of accuracy, Claude 2.1 boasts twice the accuracy of Claude 2.0. The model also improves text comprehension, particularly for legal documents, financial reports, and technical specifications, reducing incorrect answers by 30%.
Claude 2.1 can integrate domain-specific databases to address potential shortcomings of general models. Anthropic mentions that if users need programming assistance, they can instruct Claude to switch to "coding mode," converting natural language into structured output. However, this feature is still in early development.
Additionally, Claude 2.1 finally introduces the widely adopted system prompt feature seen in competitors. This allows users to preset model prompts, such as instructing the model to respond in a specific role or tone.
Anthropic also introduced Claude's pricing advantage. Users can use Claude for free for chatting, while its paid API is priced at $8 (approximately 57 RMB) per million tokens for input and $24 (approximately 172 RMB) per million tokens for output. This is lower than GPT-4 Turbo's pricing of $10 (approximately 72 RMB) per million tokens for input and $30 (approximately 215 RMB) per million tokens for output.