Anthropic is investing heavily in shaping how governments regulate artificial intelligence, pledging another $20 million to Public First Action, a nonpartisan policy organization focused on AI safeguards. This brings the AI company's total commitment to the group to $40 million, signaling that Anthropic views policy advocacy as central to its mission alongside building AI models like Claude.

The donation comes as AI capabilities are advancing rapidly. Anthropic's Claude Mythos Preview model discovered thousands of high-severity software vulnerabilities earlier this year, including flaws in every major operating system and browser. Rather than publicize these findings broadly, Anthropic released the model to a limited group of cybersecurity experts through Project Glasswing, allowing defenders to patch vulnerabilities before malicious actors could exploit them.

This scenario illustrates why Anthropic believes regulation matters. In the wrong hands, advanced AI models could threaten critical infrastructure like hospitals and power grids. The company argues that while the benefits of more capable AI are substantial, governments need time to build regulatory capacity before those risks materialize.

What Does Anthropic Want From AI Policy?

Anthropic has outlined specific policy recommendations in its Advanced AI Framework, which the company describes as "the strongest policy proposal from any frontier lab or policymaker to date." The framework calls for several concrete measures:

• Government Verification: Regulators should be able to verify that companies' safety claims are accurate and that they're following safe practices.
• Civil Penalties: Governments should have enforcement mechanisms to penalize companies that fail to meet safety standards.
• Deployment Authority: Authorities should have the power to slow or block deployment of AI models that pose a serious risk of catastrophic harm.
• Mandatory Testing: Frontier AI developers should test models that could cause catastrophic damage and submit findings to independent evaluators.
• Security Requirements: Companies should maintain robust security programs to prevent unauthorized access to their most powerful models.

Anthropic also emphasizes national security concerns. The company supports tightening export controls on advanced chips and semiconductor manufacturing equipment, as well as measures to prevent illicit access to and copying of AI models. The reasoning is straightforward: democracies currently lead in AI capability, but that advantage is fragile and could be lost if advanced technology leaks to authoritarian regimes.

How Is Anthropic's Claude Evolving Alongside These Policy Efforts?

While Anthropic advocates for regulation, it's also advancing its own AI models. The company released Claude Sonnet 5 on June 30, 2026, and immediately made it the default model across all Claude.ai tiers, including the free plan. This represents a significant shift; previously, the most capable Sonnet model was reserved for paid subscribers.

Claude Sonnet 5 introduces meaningful improvements in what AI researchers call "agentic" capabilities, meaning the model's ability to work autonomously on complex, multi-step tasks. The improvements include better handling of tool chains, where the model needs to use multiple software functions in sequence. Sonnet 5 can now dispatch multiple tool calls simultaneously rather than waiting for each one to complete, reducing delays in workflows that previously bottlenecked on sequential execution.

The model also shows significant progress on code generation. On SWE-bench Verified, a benchmark that measures real-world coding ability by having models fix actual bugs in real GitHub repositories, Sonnet 5 scored 62.4%, up from Sonnet 3.7's 49.3%. This gap matters because SWE-bench tests practical ability in complex codebases, not just algorithmic puzzle-solving.

Sonnet 5 maintains a 200,000-token context window, meaning it can process roughly 100,000 words at once. This is sufficient for most professional tasks, including analyzing entire legal documents, large codebases, or comprehensive technical specifications. The model also supports real-time web search, Python code execution in a sandboxed environment, document uploads, and desktop automation through screenshot observation and click actions.

Why Does Anthropic Make Its Best Model Free?

Deploying Sonnet 5 as the free-tier default is unusual in the AI industry. OpenAI, by contrast, keeps its most capable model, GPT-5.6, behind paid tiers. Anthropic's decision signals a philosophical stance: frontier AI capability should be broadly accessible rather than gatekept entirely behind paywalls. This approach also demonstrates confidence in the model's quality and Anthropic's ability to serve a larger user base economically.

For developers building autonomous AI workflows, Sonnet 5's improvements to tool use and error recovery address common failure modes in production systems. For researchers and knowledge workers, the combination of a large context window and improved reasoning about relationships between distant sections of documents makes it a legitimate research assistant. For enterprise teams, the integrated capabilities reduce the complexity of building AI systems that need to search the web, run code, and interact with desktop applications.

Anthropic's dual strategy, combining policy advocacy with rapid model improvement, reflects the company's view that AI development and governance must advance together. As capabilities accelerate, the stakes of getting regulation right grow higher, which is why Anthropic is investing in both building safer models and shaping the policy frameworks that will govern their use.