Anthropic Unveils Enhanced Version of Its Entry-Level Language Model

Anthropic, the AI startup founded by former OpenAI executives, has launched an updated version of its text-generating model, Claude Instant. The newest iteration, Claude Instant 1.2, offers enhanced performance through API access, delivering improved speed and cost-effectiveness.

According to Anthropic, Claude Instant 1.2 leverages the strengths of its flagship model, Claude 2, achieving impressive advancements in critical areas like mathematics, coding, reasoning, and safety. In internal assessments, Claude Instant 1.2 achieved a score of 58.7% on a coding benchmark, surpassing Claude Instant 1.1’s score of 52.8%. For math-related questions, it recorded 86.7% compared to Claude Instant 1.1's 80.9%.

“Claude Instant produces longer, more organized responses and adheres better to formatting instructions,” Anthropic stated in a blog post. “This latest version also shows enhancements in quote extraction, multilingual capabilities, and question answering.”

Furthermore, Anthropic claims that Claude Instant 1.2 reduces the likelihood of generating incorrect information, or “hallucinations,” and is more resilient against jailbreaking attempts—a method that uses intricate prompts to circumvent the safety mechanisms designed to protect large language models.

Claude Instant 1.2 also boasts a context window equal in size to Claude 2’s, allowing for 100,000 tokens. This context window refers to the text the model uses before generating new content, with tokens representing raw text fragments (e.g., the word “fantastic” is split into the tokens “fan,” “tas,” and “tic”). Together, Claude Instant 1.2 and Claude 2 can process approximately 75,000 words—around the length of “The Great Gatsby.” Models with expansive context windows tend to have better memory of recent conversations.

As we previously reported, Anthropic aims to develop a “next-generation algorithm for AI self-teaching,” as laid out in an investor pitch deck. This technology could lead to the creation of advanced virtual assistants capable of composing emails, conducting research, and producing creative works like art and literature—features we’ve started to see with existing models such as GPT-4.

However, Claude Instant is not this self-teaching algorithm; rather, it seeks to compete with entry-level offerings from OpenAI and startups like Cohere and AI21 Labs, which are also crafting and commercializing their text- and image-generating AI systems.

Since its inception in 2021, Anthropic, led by former OpenAI VP of research Dario Amodei, has raised $1.45 billion and is currently valued in the single-digit billions. While this funding is substantial, it falls short of the $5 billion the company estimates it will require over the next two years to develop its envisioned chatbot.

Anthropic claims to have “thousands” of clients and partnerships, including Quora, which offers access to Claude and Claude Instant via its subscription-based generative AI app, Poe. Additionally, Claude powers the newly launched DuckAssist tool from DuckDuckGo, providing direct answers to simple search queries alongside OpenAI’s ChatGPT. On Notion, Claude serves as a backend component for Notion AI, an integrated AI writing assistant within the Notion workspace.