Anthropic Opens Up Its Most Advanced AI Model to the Public, With Safeguards

Anthropic has released a new version of its Claude AI system to the general public, but with restrictions on how it can be used in sensitive areas like cybersecurity and weapons research. The model, called Claude 3.5 Sonnet, is the first from the company’s latest lineup to be widely available. Previously, access was limited to a small group of partner organizations due to concerns about the technology’s potential misuse.

Anthropic says Claude 3.5 Sonnet is useful for writing and debugging code, answering complex research questions, and analyzing images. At the same time, the company is offering a more powerful version, Claude 3.5 Opus, to select partners—including cybersecurity firms in its trusted testing program. That group has grown to about 200 organizations across more than 15 countries.

The company has restricted the full model over fears it could be used to identify vulnerabilities in critical infrastructure like banking systems and power grids. Critics have accused Anthropic of exaggerating these risks, but testers have since confirmed the model’s capabilities. The U.S. government, which had earlier been in a legal dispute with Anthropic over AI safety, has also tested the system.

To prevent misuse, most queries about cybersecurity, biology, or chemistry are automatically rerouted to a less capable model, Claude 3.5 Haiku, which was released in late May. Anthropic says it has also detected large-scale attempts to extract its technology for training competing AI models, and those queries are also diverted.

The company hired outside experts to spend over 1,000 hours trying to bypass these restrictions—a process known as red-teaming—and ran a bug bounty program. No one succeeded in fully unlocking the model.

Claude 3.5 Sonnet is priced at $3 per million input tokens and $15 per million output tokens—about double the cost of the Haiku model. An intensive coding session can burn through a million tokens in hours. Despite rapid revenue growth, Anthropic is not yet profitable, partly due to the high cost of computing power.

The launch comes as both Anthropic and rival OpenAI have filed for IPOs, signaling the intense financial interest in AI.