Anthropic, PBC is an American artificial intelligence (AI) public benefit corporation headquartered in San Francisco, California. Its flagship product is Claude, a series of proprietary large language models (LLMs). Anthropic was founded in 2021 with the goal of promoting AI safety by former members of OpenAI, including siblings Daniela Amodei and Dario Amodei, who are president and CEO, respectively. The company is privately held but reportedly plans an initial public offering in 2026. Valued at US$965 billion in a May 2026 Series H funding round, Anthropic is one of the world's most valuable AI pure play companies, rivaled by OpenAI.
Anthropic develops the LLM Claude in four tiers: Haiku, Sonnet, Opus, and Fable. Claude is accessible via an online chatbot, an API, and Anthropic's agent harnesses: Claude Code, Cowork, Design, and Science. Its model class Claude Mythos is available only to approved US organizations with a cybersecurity and life sciences focus. Anthropic's Project Panama, internally described as an "effort to destructively scan all the books in the world", provides training data for Claude. Before Project Panama, Anthropic downloaded pirated books and was sued, settling in 2025 a copyright infringement lawsuit from authors for $1.5 billion.
Anthropic conducts LLM research including on mechanistic interpretability, safety, alignment, and societal impact. Notable researchers working at Anthropic include Andrej Karpathy, John Jumper, Jan Leike, Chris Olah and Amanda Askell.
Anthropic partners with Palantir to provide Claude to US federal agencies including the Department of Defense (DoD) and the Intelligence Community; Claude was used in the 2026 Iran war and the 2026 United States intervention in Venezuela. In February 2026, the DoD demanded the ability to use Claude for all lawful purposes, but Anthropic refused to drop restrictions against mass domestic surveillance and fully autonomous weapons. The dispute led the Trump administration to phase out Anthropic products from government use, a decision blocked by a preliminary injunction from a federal judge who called it "First Amendment retaliation". Throughout 2026, Anthropic accused Chinese competitors including DeepSeek, Moonshot AI, Alibaba Cloud, and Z.ai of distilling Claude for their models.
2024: Claude 3 and Claude 3.5 Releases
The year 2024 marked a significant expansion of Anthropic's Claude model family. In March 2024, the company released the Claude 3 series, introducing three models: Haiku, Sonnet, and Opus. These models demonstrated improvements in reasoning, coding, and multilingual capabilities compared to previous versions. Claude 3 Opus, the largest model, achieved state-of-the-art results on several benchmarks, including undergraduate-level knowledge and graduate-level reasoning tests. The Claude 3 family also introduced enhanced vision capabilities, allowing the models to process images and documents.
In June 2024, Anthropic released Claude 3.5 Sonnet, a mid-tier model that outperformed the larger Claude 3 Opus on many tasks, including coding and mathematical reasoning. This was followed by Claude 3.5 Haiku in October 2024, which offered a faster and more cost-effective option. The Claude 3.5 models introduced a feature called "computer use", enabling the AI to interact with graphical user interfaces by moving a cursor and clicking buttons, a step toward autonomous agent behavior. These releases positioned Anthropic as a leading competitor in the generative AI space, challenging models from OpenAI and Google DeepMind.
Model Architecture and Training
Claude models are built on the transformer architecture, a type of neural network that has become the standard for deep learning in natural language processing. Anthropic employs a technique called Reinforcement Learning from AI Feedback (RLAIF) to align models with human preferences, which involves using AI-generated comparisons rather than purely human feedback. This approach is part of Anthropic's broader focus on AI safety and machine learning research.
The training process for Claude involves large-scale data augmentation and the use of learning rate schedules to optimize performance. Anthropic has also contributed to research on mechanistic interpretability, aiming to understand the internal workings of neural networks. The company's models are trained on a diverse corpus of text, including books, articles, and web content, though the specifics of the training data are not fully disclosed.
Safety and Alignment Research
Anthropic's mission centers on ensuring that AI systems are safe and aligned with human values. The company conducts research on model pruning, gradient clipping, and other techniques to improve model robustness. Its safety team, led by researchers such as Jan Leike and Chris Olah, focuses on understanding and mitigating risks associated with advanced AI, including potential misuse and unintended behaviors.
In 2024, Anthropic published several papers on topics like multi-head attention analysis and cross-attention mechanisms, contributing to the broader field of artificial intelligence research. The company also advocates for responsible AI development, participating in policy discussions and collaborating with other organizations like the Berkeley AI Research lab.
Partnerships and Compute Infrastructure
To support its large-scale model training, Anthropic has secured substantial computing resources. In October 2025, the company announced a cloud partnership with Google, gaining access to up to one million of Google's custom Tensor Processing Units (TPUs), potentially adding more than one gigawatt of AI compute capacity by 2026. This partnership complements Anthropic's earlier agreements with Amazon Web Services and Microsoft Azure. In November 2025, Nvidia and Microsoft were expected to invest up to $15 billion in Anthropic, and Anthropic said it would buy $30 billion of computing capacity from Microsoft Azure running on Nvidia AI systems.
Anthropic also collaborates with chipmakers like AMD and Intel to optimize Claude for various hardware platforms. The company's reliance on TSMC for chip fabrication and Broadcom for networking components highlights its integration into the broader semiconductor supply chain.
Commercial and Government Applications
Anthropic's Claude models are used across various industries, including finance, healthcare, and technology. The company offers an API that developers can integrate into their applications, as well as specialized tools like Claude Code for coding assistance. In July 2025, Anthropic signed a two-year contract with the US Department of Defense for $200 million, leading to the integration of Claude with the Pentagon's classified intelligence networks and Palantir's data platform. This partnership has been controversial, given Anthropic's stated safety principles, but the company has maintained that it supports lawful uses of AI in national security contexts.
In September 2025, Anthropic announced that it would stop selling its products to groups majority-owned by Chinese, Russian, Iranian, or North Korean entities due to national security concerns. This decision reflects the growing geopolitical tensions around AI technology.
Controversies and Legal Issues
Anthropic has faced legal challenges related to its training data. In 2025, the company settled a copyright infringement lawsuit from authors for $1.5 billion, stemming from its use of pirated books before Project Panama. The settlement highlighted the ethical and legal complexities of training large language models on copyrighted material.
In 2026, Anthropic accused Chinese competitors, including DeepSeek, Moonshot AI, and Alibaba Cloud, of knowledge distillation "attacks" that violated its terms of service. The company alleged that these firms generated over 16 million chats with Claude using around 24,000 fraudulent accounts, which were then used as training data. These accusations reflect the competitive and often contentious nature of the AI industry.
Future Outlook
As of 2026, Anthropic continues to expand its model capabilities and commercial reach. The company's reported plans for an initial public offering in 2026, coupled with its high valuation, indicate strong investor confidence. However, its disputes with the US government and ongoing legal battles over data usage pose potential risks. Anthropic remains committed to its mission of AI safety, even as it navigates the complex landscape of commercial AI development.