Yes, Claude uses industry-standard best practices for data handling and retention. See our Privacy Policy for more details. All commercial deployments are covered by Anthropic's Data Protection Addendum, which is available upon request. 

Please find our model card for Claude 2.0 and Claude 2.1 here. 

Our API is designed to be a backend that incorporates Claude into any application you’ve developed. Your application sends text to our API, then receives a response via server-sent events, a streaming protocol for the web. We have API documentation with drop-in example code in Python and Typescript to get you started.

We currently offer two versions of Claude.

Claude - our most powerful model, which excels at a wide range of tasks from sophisticated dialogue and creative content generation to detailed instruction following.

Claude Instant - a faster and cheaper yet still very capable model, which can handle a range of tasks including casual dialogue, text analysis, summarization, and document question-answering.

You can find our model pricing here.

Claude has the most training in English, but also performs well in a range of other common languages, and still has some ability to communicate in less common languages. Claude also has extensive knowledge of common programming languages!

No. Claude is designed to be self-contained, and will respond without searching the internet. You can, however, provide Claude with text from the internet and ask it to perform tasks with that content.

Constitutional training is a process for training a model to adhere to a “constitution” of desired behavior. The core Anthropic model has been finetuned with constitutional training with the goal of becoming helpful, honest, and harmless. You can learn more about constitutional training here. 

Helpful, Honest, and Harmless (HHH) are three components of building AI systems (like Claude) that are aligned with people’s interests.

- Helpful: Claude wants to help the user

- Honest: Claude shares information it believes to be true, and avoids made-up information

- Harmless: Claude will not cooperate in aiding the user in harmful activities

While no existing model is close to perfection on HHH, we are pushing the research frontier in this area and expect to continue to improve. For more information about how we evaluate HHH in our models, you can read our paper here.

Claude’s behavior can be extensively modified using prompting. Prompts can be used to explain the desired role, task, and background knowledge, as well as a few examples of desired responses. You can read more about prompt design here. 

In the vast majority of cases, we believe well-crafted prompts will get you the results you want without the expense or delay of fine-tuning. However, some large enterprise users may significantly benefit from fine-tuned models. Please contact us to discuss whether your needs might be best addressed with a fine-tuned model by filling out this form. 

The combined context window for input and output is about 200,000 tokens, which works out to roughly 150,000 words, depending on the type of content.

Not at this time! We find the open source SBERT embeddings to be good enough for most purposes.