Claude

Claude is an artificial intelligence (AI) assistant developed by Anthropic, accessible via API and a chat interface on the company's website. Claude is capable of a wide variety of conversational and text-processing tasks, including summarization, search, creative and collaborative writing, Q&A, coding, general open-ended conversation, outlining, and providing advice on a broad range of subjects. Claude is particularly well-suited to supporting creative or literary use cases, with the ability to take direction on tone or "personality." Trained using constitutional AI, a technique developed by Anthropic that better aligns AI systems with human intentions, customer reports suggest Claude is less likely to produce harmful outputs.

The company worked with key partners, such as Notion, Quora, and DuckDuckGo, testing the AI assistant while in closed alpha. Anthropic imposed an embargo on media coverage of Claude that was lifted in January 2023. The model became widely available to companies requesting access on March 14, 2023. Since the initial version, Anthropic has released two major updated versions of the model—Claude 2 (July 11, 2023) and Claude 2.1 (November 21, 2023).

All versions of Claude are general-purpose large language models that use a transformer architecture and are trained via unsupervised learning, reinforcement learning from human feedback (RLHF), and Constitutional AI (including both a supervised and Reinforcement Learning phase). Claude is trained on a proprietary mix of publicly available information from the Internet, datasets, licenses from third-party businesses, and data Anthropic users affirmatively share or that crowd workers provide. Anthropic has made some of the human feedback data used to finetune Claude public as well as the company's RLHF and red-teaming research. Claude is self-contained, meaning it responds without searching the internet. Its training data stops in early 2023, but users can provide the assistant with text from the internet to perform tasks.

Roughly 90 percent of Claude's training data is English, but the model can also perform tasks in a range of other common languages with some abilities to communicate in less common languages. Claude has extensive knowledge of common programming languages. The model's behavior can be modified via prompting, and it is possible to fine-tune the model, although Anthropic states this is only really beneficial for large enterprise users. Claude's combined context window for input and output is roughly 100,000 tokens (~70,000 words), depending on the type of content. The model does not support embeddings, instead using the open-source SBERT embeddings.

Anthropic offers a full and instant version of Claude:

• Claude—the full version of the assistant that excels at a wide range of tasks, including sophisticated dialogue, creative content, and detailed instruction.
• Claude Instant—a faster and cheaper model with reduced latency. Claude instant remains capable of a range of tasks, including casual dialogue, text analysis, summarization, and document comprehension.

Claude 2 replaced the first version of the model. It is an evolution of the original release with a series of small but meaningful improvements that build on the company's research into reliable, steerable, and interpretable AI systems. Claude 2.1 offers the same performance as Claude 2 but with a number of advancements, including the following:

• 200k token context window (double the previous models)
• A significant reduction in model hallucination rates
• System prompts that allow users to provide custom instructions to Claude for better performance
• A new beta feature called tool use

Tool use means users can integrate Claude with existing processes, products, and APIs. Claude 2.1 can now orchestrate developer-defined functions or APIs, search web sources, and retrieve information from private databases. Users can define a set of tools, then let Claude decide which one is required to complete a specific request.

After its release, Claude 2.1 became available through Anthropic's API and chat interface for both free and pro tiers. However, the 200k token context window is only available for Claude Pro users.

Claude 2 scored 76.5% on the multiple-choice section of the Bar exam, an increase from 73.0% with Claude 1.3. Compared to college students applying to graduate school, Claude 2 scores above the 90th percentile on the GRE reading and writing exams. On the Codex HumanEval, a Python coding test, Claude 2 scored 71.2%, and on GSM8k, a large set of grade-school math problems, Claude 2 scored 88%. Using Anthropic's internal red-teaming evaluation, Claude 2 was 2x better at giving harmless responses compared to Claude 1.3.

Claude 2.1 offers the same performance as Claude but with a significant reduction in rates of hallucination. This includes a 2x decrease in false statements compared to Claude 2.0. This result was shown by testing both models using a curated set of complex, factual questions designed to probe known weaknesses in the models. Claude 2.2 also shows improvements in comprehension and summarization that demand a high degree of accuracy. Anthropic evaluations show Claude 2.1 demonstrated a 30% reduction in incorrect answers and a 3-4x lower rate of mistakenly concluding a document supports a particular claim.

After the release of Claude 2.1, Anthropic updated its pricing for its three available models (Claude Instant, Claude 2, and Claude 2.1). Prices are based per million prompt and completion tokens:

• Claude Instant (100k context window)—Prompt $1.63/million tokens, completion $5.51/million tokens
• Claude 2 (100k context window)—Prompt $8/million tokens, Completion $24/million tokens
• Claude 2.1 (200k context window)—Prompt $8/million tokens, Completion $24/million tokens

Anthropic is working with thousands of businesses using the Claude API. Examples include those below:

• Jasper—a generative AI platform for enabling users to scale their content strategies.
• Sourcegraph—a code AI platform helping customers write, fix, and maintain code. Sourcegraph's coding assistant "Cody" utilizes Claude 2's reasoning abilities to provide accurate answers.