Anthropic raises $450 million to expand Claude and its safety

Anthropic announced Tuesday that it had raised a $450 million funding round led by Spark Capital, with participation from Google, Salesforce Ventures, Sound Ventures and Zoom Ventures. The money will go toward expanding Claude, its AI-powered assistant, and funding research into the safety of advanced language models.

The deal puts more resources behind one of the young companies competing directly with OpenAI. Anthropic was founded in 2021 by former OpenAI employees, including siblings Dario and Daniela Amodei, and has built its identity around a difficult problem: creating powerful models without losing the ability to control them.

Claude enters the assistant race

Claude is a language model: a system trained on large amounts of text to answer questions, summarize documents, write and hold conversations. Anthropic introduced it in March and has made it available to a limited group of companies through its API, the way businesses integrate the model into their own products. It can also be tried through Poe, Quora’s assistant service.

The company announced this month that Claude can process up to 100,000 tokens of context. In simple terms, that means it can handle very long documents or extended conversations without losing track of what was said earlier. The capability is relevant to tasks such as reviewing contracts, analyzing reports, searching document repositories and assisting teams that work with large volumes of information.

OpenAI had raised the bar with GPT-4, released in March, while Google had just opened Bard to the public in the United States and the United Kingdom. The funding does not guarantee that Anthropic will catch up with those companies in users or distribution, but it gives the startup room to compete in a field where training and operating large models requires extremely expensive computing infrastructure.

Safety as both a product and a scientific bet

Anthropic’s main point of differentiation is its work on alignment, the field focused on making AI follow the intentions and constraints set by people. Its best-known technique is Constitutional AI: rather than relying solely on human evaluators to score responses, the model learns to review itself with help from an explicit set of principles.

Those principles function like a constitution: the system should favor responses that are useful, honest and harmless, and reject requests that could cause harm. The company has said the method combines human oversight with AI-generated feedback. It does not eliminate errors or guarantee that Claude is immune to malicious instructions, but it aims to make the reasoning behind the assistant’s refusal or reformulation of a request more visible.

The issue matters because conversational assistants are no longer simple demonstrations. They are being integrated into search engines, programming tools, customer service and office software. A failure can result in anything from an invented answer to the leakage of sensitive information or the automation of harmful content. Safety, therefore, is starting to become a commercial requirement for selling these systems to businesses, as well as a research problem.

Google strengthens a second generative-model bet

Google had already invested $300 million in Anthropic in February, in addition to taking an approximately 10% stake and reaching an agreement to use its cloud computing infrastructure. Its presence in the new round confirms that the technology giant wants to maintain a significant position in an independent lab while developing Bard and its own models.

For Google, backing Anthropic offers access to safety research and talent without fully absorbing the company. For Anthropic, it means access to computing capacity on Google Cloud, a decisive resource when training a frontier model can require thousands of specialized accelerators for weeks or months.

The round brings the total capital raised by Anthropic since its founding to nearly $1 billion. Its next test will be less financial than technical: proving that Claude can deliver competitive responses against GPT-4 and other assistants, maintain that quality as it reaches more users, and turn its safety mechanisms into a verifiable advantage for real customers.