May 2025 Summaries

Anthropic appoints Reed Hastings to its board of directors, bringing extensive experience from founding Netflix and serving on the boards of Facebook, Microsoft, and Bloomberg. Hastings' leadership experience, deep philanthropic work, and commitment to addressing AI's societal challenges make him uniquely qualified to guide Anthropic at this critical juncture in AI development. His recent focus on AI and Humanity through his research initiative at Bowdoin College aligns closely with Anthropic's own research priorities. Hastings has also spent decades working to expand opportunity in education and global development, both domestically and internationally, making him an ideal addition to the board as they continue building AI that helps rather than harms.

The Anthropic API has introduced four new capabilities that enable developers to build more powerful AI agents: a code execution tool, an MCP connector, a Files API, and extended prompt caching. The code execution tool allows Claude Opus 4 to run Python code in a sandboxed environment for data analysis, while the MCP connector connects Claude to external systems without requiring custom infrastructure. The Files API simplifies file uploads and access, and extended prompt caching enables agents to maintain context over extended periods with reduced costs. These new capabilities join existing features like web search and citations as part of a comprehensive toolkit for building AI agents.

The latest models from Anthropic, Claude Opus 4 and Claude Sonnet 4, set new standards for coding, advanced reasoning, and AI agents. These hybrid models offer two modes: near-instant responses and extended thinking for deeper reasoning. They are available on the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI, with pricing consistent with previous Opus and Sonnet models. Claude Opus 4 is the world's best coding model, delivering sustained performance on complex tasks, while Claude Sonnet 4 significantly improves on Sonnet 3.7's industry-leading capabilities. The new models come with extended thinking capabilities, tool use, parallel execution, memory improvements, and reduced shortcuts or loopholes. Additionally, Claude Code, a collaborative development platform, is now generally available, allowing developers to collaborate with Claude in their preferred IDEs and tools. The models have been extensively tested and evaluated for AI safety levels, including ASL-3.

We have activated the AI Safety Level 3 (ASL-3) Deployment and Security Standards in conjunction with launching Claude Opus 4, a newly released model. The ASL-3 measures are designed to increase internal security and limit the risk of misuse specifically for the development or acquisition of chemical, biological, radiological, and nuclear (CBRN) weapons. These measures make it harder to steal model weights and cover a narrowly targeted set of deployment measures that limit the risk of Claude being misused for CBRN-related tasks. The new measures are part of Anthropic's Responsible Scaling Policy, which aims to increase increasingly capable AI models warranting stronger deployment and security protections. The ASL-3 Standard involves constitutional classifiers that monitor model inputs and outputs and intervene to block harmful CBRN information. Ongoing refinement and iteration will be necessary to improve the effectiveness of these measures and address potential issues.

We are launching a new bug bounty program to stress-test our latest safety measures, focusing on universal jailbreaks in Constitutional Classifiers system. This initiative tests an updated version of the system, which follows principles that define what type of content should and shouldn’t be allowed when interacting with Claude. The program is invite-only and offers rewards up to $25,000 for verified universal jailbreaks found on the unreleased system. Participants will receive early access to test our classifiers on Claude 3.7 Sonnet, and the program aims to contribute to the work done over several months to iterate and stress-test ASL-3 safeguards.

With the introduction of web search on the Anthropic API, developers can now build AI applications and agents that tap into current information from across the web without needing to manage their own web search infrastructure. Claude, a comprehensive knowledge model, uses its reasoning capabilities to determine whether searching the web would be beneficial for providing more accurate responses, generating targeted search queries, retrieving relevant results, and analyzing them for key information. This enables developers to power a wide range of use cases across various industries, such as financial services, legal research, developer tools, and productivity, with real-time data and specialized knowledge. Additionally, web search includes citations to source materials, ensuring accuracy and accountability, and provides admin settings to control domain allow lists, block lists, and organization-level management. The feature is also available in Claude Code, enhancing development workflows with current API documentation and technical articles.

Anthropic's AI for Science program aims to accelerate scientific research and discovery by providing free API credits to support researchers working on high-impact projects, particularly in biology and life sciences. The initiative aligns with Anthropic's vision of building AI systems that bring value to humanity and supports applications where AI can assist in accelerating processes related to understanding complex biological systems, analyzing genetic data, and increasing agricultural productivity. Researchers interested in the program can apply through a designated application form, which will be reviewed by Anthropic's team, including subject matter experts in relevant fields. The program is designed to reduce the time and resources needed for scientific discovery, helping to address some of humanity's most pressing challenges.

Claude can now connect to a wide range of tools and apps through its Integrations feature, allowing users to access deep context about their work and take actions across multiple surfaces. This expansion enables Claude to become a more informed collaborator, helping users execute complex projects with expert assistance at every step. Additionally, Claude's Research capabilities have been enhanced with an advanced mode that searches the web, Google Workspace, and now Integrations too, delivering comprehensive reports complete with citations in up to 45 minutes. With Integrations, users can discover and connect any number of tools to Claude, expanding its capabilities and functionality.