Anthropic says its Claude AI models are taking a larger role in the company’s own artificial intelligence research and development work. Claude now leads about 26% of Anthropic’s AI research and development efforts, up from about 1% in March, according to research cited by Reuters. Anthropic said the systems remain under human supervision and do not operate independently. More than 90% of the company’s research work involved some form of human-AI collaboration in August.
Anthropic has also opened applications for its new Life Sciences Verification Program after onboarding dozens of organizations through an early access program. This program is initially limited to teams and institutions, but will later include individual users on Pro and Max plans. The program is designed to give verified life-sciences researchers broader access to Claude for work including drug discovery, biology research, clinical development and manufacturing.
Each applicant must go through a verification process that checks research credentials, security standards, and ethical oversight practices. Standard Use grants are available for most life science projects, including the majority of biology research workflows. These grants currently cover Claude Mythos 5.1, Opus 5 and Sonnet 5, as well as future models when they are released. Standard grants can cover entire teams and are renewed annually.
High-risk Use grants are an add-on option for projects that cannot proceed under Standard Use rules. These grants remove safeguards that specifically block life-sciences requests and apply to a single project rather than an entire team. Other safeguards, including cyber protections, remain in place. They must be renewed every six months and are available for Claude Opus 5 and Claude Sonnet 5 at launch.
Anthropic is also working with the U.S. government to make these grants more broadly available for its Claude Mythos model. High-risk access to Mythos remains limited to a small group of organizations that receive additional vetting. The company has developed specific safeguards against three major threat models: access compromise, insider threats, and agent misuse. These protections were designed with input from enterprise security leaders and tie each organization’s access to the use cases listed in its application.
Anthropic monitors traffic related to the program to detect activity outside approved scopes. When unusual behavior is found, admins are notified and must respond within a set timeframe. The system shifts enforcement from real-time blocking to offline monitoring of usage patterns. Anthropic says this makes it easier to detect suspicious behavior spread across multiple requests or sessions.
This approach allows legitimate work to proceed with fewer interruptions but requires storing data for 30 days. The data collected is compartmentalized and cannot be used for training models or accessed by Anthropic’s life sciences research team.
The safeguards come as Anthropic reports attempts to misuse Claude for biological research that could support harmful activity. In a report covering activity from December 2025 to August 2026, Anthropic described five case studies involving biological misuse. The cases included attempts involving research on viruses and other work that Anthropic believed could contribute to biological weapons development. Anthropic said it disrupted the activity and used information from the cases to strengthen its safeguards.
The life-sciences expansion is taking place as Anthropic prepares for a possible initial public offering. The company confidentially filed for a U.S. IPO in June. Reuters reported in September that an offering could value Anthropic at about $2 trillion and raise as much as $100 billion. Nvidia is discussing a possible investment of up to $10 billion as part of the offering, though no final agreement has been announced.
Annualized revenue reached more than $65 billion by the end of July, a sharp increase from about $9 billion at the end of 2025. The figure is a revenue run rate rather than reported annual revenue. Anthropic has benefited from growing enterprise use of Claude and major partnerships with technology companies including Amazon and Google.
Anthropic has also told investors that it expects positive adjusted operating income for a second consecutive quarter, according to a Financial Times report cited by Reuters. That measure excludes some major costs and should not be treated as the same thing as GAAP profitability. Anthropic has not yet published full financial statements through a public IPO filing.
CEO Dario Amodei has raised concerns about AI’s potential for causing catastrophic harm, including existential risks. He has spoken publicly about risks involving cyberattacks, biological weapons and AI systems acting outside human control. In a June policy essay, Amodei called for mandatory testing of advanced AI models by qualified independent organizations. He compared the regulatory approach he favors to oversight used in industries such as aviation and pharmaceuticals.
Former Anthropic researcher Jacob Coxon also raised concerns about rapid AI development when he recently left the company. Reuters reported that Coxon warned that advanced AI could pose severe risks to humanity within the next decade. His statements represent his own assessment rather than an established probability of what will happen.
Anthropic operates as a public-benefit corporation. Under its corporate structure, directors are permitted to consider the company’s stated public-benefit mission alongside shareholder interests. Anthropic says its stated purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity.
The company faces competition from OpenAI and lower-cost models developed in China and elsewhere. OpenAI released GPT-6 Astra in September and describes it as its most capable model for complex professional work, coding and computer use. Benchmark results vary by test, so broad claims that one model is better than another depend on the task and evaluation being used.
Anthropic’s possible $2 trillion IPO valuation would equal about 31 times its current $65 billion annualized revenue run rate. That valuation is still a reported target rather than the company’s current public-market value. Investors will receive more detailed information about revenue, costs and risks when Anthropic makes its IPO registration documents public.
Sam Bankman-Fried and the collapsed cryptocurrency exchange FTX were early Anthropic investors. FTX invested about $500 million in Anthropic. A bankruptcy judge later approved the sale of FTX’s Anthropic stake so the estate could raise money to repay creditors. FTX held about 7.84% of Anthropic at the time approval for the sale was granted in 2024. Bankman-Fried himself no longer has an ownership stake through FTX.
Dustin Moskovitz, a Facebook co-founder and major philanthropist, also has a minority ownership interest in Anthropic. Coefficient Giving, formerly known as Open Philanthropy, has provided funding to the Tarbell Center for AI Journalism. Tarbell says its donors have no editorial control over its reporting, fellowships or grants.
Anthropic has also launched Claude Corps, a program that plans to train and place 1,000 fellows with at least 400 nonprofit organizations. The fellows are not Anthropic employees. They are employed by CodePath while Anthropic funds the program and provides Claude training and technical support. The first group of about 100 fellows is scheduled to begin in October 2026.
Amodei published a policy essay in June seeking mandatory third-party testing of frontier models. More recently, he has called for slowing the pace of frontier AI development while stronger safeguards are put in place. Other technology executives and policymakers disagree on how much regulation is needed and whether slowing development could weaken the United States in its competition with China.
President Donald Trump has taken a generally less restrictive approach to AI regulation. He signed an executive order in December 2025 directing the federal government to challenge certain state AI laws that the administration considers overly burdensome. The order does not automatically eliminate all state AI laws and specifically leaves some areas available for state regulation.
Trump has also dismissed some recent warnings about catastrophic AI risks as a “hoax.” At the same time, his administration issued a separate June 2026 order calling for cooperation with AI companies on voluntary cybersecurity testing of advanced models. The competing positions show that federal AI policy is still developing as companies including Anthropic push for stronger testing and oversight of the most capable systems.