Anthropic

Anthropic is an American artificial intelligence company headquartered in San Francisco that develops frontier AI systems with an institutional focus on safety, interpretability, steerability, and long-term benefit to humanity. It was incorporated in 2021 as a public benefit corporation by seven co-founders who had departed OpenAI between late 2020 and January 2021 after internal disagreements over the prioritization of existential safety research relative to commercial productization accelerated by the Microsoft partnership. The name Anthropic references the anthropic principle from cosmology to indicate the goal of producing AI aligned with the conditions for human flourishing.

Anthropic operates under a governance model that includes a Board of Directors and an independent Long-Term Benefit Trust empowered to elect directors and safeguard the public benefit purpose against short-term pressures. It codifies seven operating values be good ancestors, seek the truth, be courageous, be collaborative, practice humility, uphold integrity, and deliver excellence to shape decisions from daily operations to scaling choices. In 2022 the company introduced Constitutional AI, a technique that trains models via AI feedback against an explicit set of natural-language principles rather than relying exclusively on human preference data. Anthropic uses the method to train its Claude models and published Claude's constitution in May 2023. A substantially longer revision released in January 2026 ranked being broadly safe, broadly ethical, compliant with Anthropic's guidelines, and genuinely helpful in descending order of priority.

The firm maintains a Responsible Scaling Policy that assigns models to AI Safety Levels based on evaluations of malicious use and loss-of-control risks, committing to implement layered safeguards or pause development when thresholds are not met. Its Claude family began commercial API availability on March 14, 2023 and evolved through releases that expanded context length, added multimodality, tool use, and agentic features, with progress measured by evaluations covering autonomy, task performance, research acceleration, and specialized domains such as cybersecurity vulnerability discovery.

Anthropic raised capital through successive rounds that included major investments from Amazon Web Services and Google, while forming primary compute partnerships with Amazon Web Services for dedicated Trainium and Inferentia capacity and with Google for TPU supply and equity investment supplemented by clean power agreements. In October 2026 it filed for an initial public offering of Class A common stock on the Nasdaq Global Select Market under the symbol ANT; an amended prospectus reported third-quarter 2026 revenue of $17.0 billion. The company has published research on mechanistic interpretability and scaling laws, collaborated with the National Nuclear Security Administration on nuclear risk assessment, and achieved rapid enterprise adoption with growing annualized revenue while positioning itself through transparency and safety commitments relative to competitors. In June 2025 a federal judge ruled that Anthropic's training of Claude models on legally purchased and destructively scanned print books was fair use, while the company later paid $1.5 billion to settle claims arising from its earlier use of millions of pirated digital books.

Founding and Origins

Departure from OpenAI

In late 2020 Dario Amodei resigned from his position as vice president of research at OpenAI after repeated internal debates in which he argued that the organization was deprioritizing long-term existential safety research in favor of accelerated commercial productization and revenue generation following its strategic partnership with Microsoft. Six additional senior OpenAI personnel departed with him between December 2020 and January 2021, creating a group of seven co-founders who cited fundamental disagreements over safety priorities and the company's shift from its original nonprofit charter toward rapid commercialization. The seven co-founders and their former roles at OpenAI were Dario Amodei, vice president of research; Daniela Amodei, vice president of policy and operations; Jared Kaplan, lead researcher on scaling laws for neural language models; Tom Brown, core technical contributor to the GPT-3 training run and architecture; Jack Clark, head of policy and communications; Chris Olah, distinguished researcher and leader of the mechanistic interpretability team; and Sam McCandlish, senior research scientist focused on large-scale compute-efficient training methods. These researchers collectively represented hundreds of person-years of frontier AI development experience and had been responsible for many of OpenAI's most cited technical outputs on language model scaling, alignment techniques, and interpretability between 2018 and 2020. The group maintained that OpenAI's leadership had begun to treat safety as a secondary consideration that could be addressed concurrently with deployment rather than as a prerequisite that should constrain scaling timelines, a stance that became untenable after internal forecasts indicated that models approaching GPT-4-level performance would require novel alignment approaches not then prioritized by the company. Disagreements intensified around resource allocation, with the departing team advocating for at least 30 percent of compute and personnel dedicated exclusively to alignment and safety infrastructure while OpenAI leadership favored directing the majority toward capability expansion and API commercialization. None of the seven co-founders held equity in OpenAI at the time of departure, and all signed standard post-employment agreements that permitted them to recruit additional talent after a brief cooling-off period. By February 2021 the group had coalesced around plans to establish an independent laboratory that would embed safety constraints directly into model training objectives from the outset rather than retrofitting them after capability milestones had been reached, directly addressing the philosophical and operational rifts that precipitated the 2020-2021 exodus.

Incorporation as Public Benefit Corporation

Anthropic was incorporated as a public benefit corporation in San Francisco, California. The selection of the name Anthropic was made to invoke the anthropic principle from cosmology, which considers the ways in which the universe is suited for the development of intelligent life, thereby signaling the company's intention to build AI that is correspondingly suited to the needs and long-term welfare of human beings. The binding purpose under its PBC status is explicitly focused on the long-term benefit of humanity, a legal requirement that compels the company to weigh this factor in all major decisions and prevents the subordination of safety and societal good to commercial pressures. This legal formation choice created a corporate structure in which the pursuit of long-term human benefit is not merely a value statement but a binding fiduciary duty backed by the force of law.

The PBC designation requires the company to publish regular benefit reports detailing its progress toward this purpose, adding a layer of transparency and accountability to its commitment to developing AI that supports rather than endangers the future of human society. By embedding this focus on long-term human benefit into its foundational legal documents at the moment of incorporation, the public benefit corporation structure established an enforceable corporate charter that obligates directors to balance shareholder returns against measurable contributions to humanity's sustained flourishing across extended time horizons. The certificate of incorporation specifies that the company exists to ensure artificial intelligence advances in ways that produce positive material impacts on society with particular emphasis on long-term human benefit, creating a legal standard that must be considered in product development, research prioritization, partnership selection, and resource allocation. This framework was deliberately adopted during formation to institutionalize resistance to short-term incentives that could otherwise compromise the company's orientation toward humanity's collective future.

Corporate Governance and Values

Board and Long-Term Benefit Trust

Anthropic operates as a Delaware public benefit corporation legally obligated to balance the interests of stockholders with its public benefit purpose of developing AI systems that are reliable, interpretable, steerable, and beneficial for humanity over the long term. The Board of Directors includes co-founders Dario Amodei, who serves as chairman, and Daniela Amodei along with independent directors chosen for relevant expertise. Board decisions are required to take the company's public benefit purpose into account, creating fiduciary duties that extend beyond pure profit maximization.

A central feature of the governance model is the independent Long-Term Benefit Trust, created at the time of incorporation in 2021 as a separate legal entity that holds a majority of the company's governance rights in perpetuity to protect the mission from short-term pressures. The Trust is overseen by trustees selected for their expertise in science, ethics, policy, and long-term societal risks who serve with the mandate to prioritize the public benefit purpose above other considerations including shareholder returns. Trustees are barred from holding financial stakes in Anthropic or its investors. The Trust possesses the power to elect and remove a majority of the members of the Board of Directors, ensuring that mission-aligned oversight predominates on key decisions including those involving model scaling, deployment, and safety policies. In its IPO registration statement Anthropic stated that, immediately prior to the offering, it would issue a single share of Class F common stock with no economic rights to a limited liability company of its seven co-founders, entitled to 50.1% of the total voting power on matters submitted to stockholders generally, including the election of directors other than the Class T directors elected by the Long-Term Benefit Trust, and that the company would be a controlled company under Nasdaq rules.

Major changes to the company's certificate of incorporation that would affect the Trust's authority or the public benefit purpose require approval that reflects the Trust's governing role. This governance design draws from other public benefit corporations but was tailored to the distinctive challenges of frontier AI development and its potential long-term impacts. The Trust receives information necessary to fulfill its oversight responsibilities regarding model capabilities, safety evaluations, and risk assessments. This structure has been highlighted in company disclosures as important for maintaining focus on safety and for recruiting researchers concerned with institutional safeguards against excessive commercialization. Details on the Trust's charter and governance mechanics are available through Anthropic's official corporate responsibility and benefit reporting.

On 18 October 2026 Anthropic announced the Anthropic Public Benefit Fund, a new entity to which it said it would commit a portion of its shares. Subject to stockholder approval, 5% of Anthropic's shares would be committed initially, with an additional 2% that may be added at the board's discretion for every 50% increase in the company's value above a set baseline. The Fund is established as a wholly owned LLC rather than a 501(c)(3), which Anthropic said gives flexibility between different forms of giving; Anthropic stated that none of the Anthropic equity held by the LLC will be used for lobbying or political giving. Anthropic said that, combined with its co-founders' pledge to donate 80% of their equity holdings and an employee equity-donation match program, roughly 20% of its shares would be dedicated to the public good.

Core Operating Values

Anthropic explicitly catalogs seven core operating values that are intended to guide decision-making from daily operations to strategic choices on AI development. These values are be good ancestors, seek the truth, be courageous, be collaborative, practice humility, uphold integrity, and deliver excellence. Being good ancestors involves prioritizing the long-term interests of humanity in AI-related decisions. Seeking the truth requires rigorous honesty in evaluations of model performance and potential societal impacts. Being courageous entails making difficult choices that may delay product launches or require additional investment in safety measures. Being collaborative means engaging constructively with the wider AI ecosystem to raise standards for safe development. Practicing humility means recognizing the limitations of current knowledge about advanced AI and avoiding overconfidence in any particular approach. Upholding integrity means ensuring that all actions are consistent with stated principles. Delivering excellence means insisting on the highest standards in technical and organizational work. The values are operationalized through a company-wide framework that requires all major proposals to include a section explaining their alignment with these principles.

Dario and Daniela Amodei

Dario Amodei serves as co-founder and chief executive officer of Anthropic where he directs overall company strategy with a focus on technical research priorities and frontier model development. He received a Ph.D. in physics from Stanford University and conducted early career research in machine learning systems at Google Brain prior to his current responsibilities. In this role at Anthropic Amodei spends the majority of his time collaborating directly with research teams on projects involving neural scaling behaviors capability assessments and alignment methodologies while also engaging directly with enterprise customers to understand real-world model applications. He has personally overseen the iterative development cycles for the Claude family of models emphasizing measurable improvements in reasoning coherence and task completion rates. Amodei holds specific accountability for Anthropic's relationships with major compute providers ensuring that infrastructure decisions align with projected capability roadmaps through 2030. Daniela Amodei serves as co-founder and president of Anthropic where she maintains responsibility for operational infrastructure talent acquisition policy coordination and business development functions. She holds an undergraduate degree in biochemistry and accumulated more than a decade of experience in research operations and organizational scaling within both biotechnology firms and technology startups before assuming her current position. At Anthropic she directly manages the expansion of the company's workforce while implementing structured hiring processes that prioritize candidates with combined technical and ethics backgrounds. Daniela Amodei leads weekly cross-functional synchronization meetings that integrate feedback from research policy and commercial teams and she holds final approval authority on all external partnership agreements exceeding $100 million in committed value. The siblings maintain distinct portfolios with Dario Amodei concentrating on scientific output metrics and Daniela Amodei focusing on execution metrics. They conduct joint quarterly reviews of company-wide key performance indicators that encompass both capability progress and operational sustainability measures without overlap in day-to-day decision ownership. Dario Amodei has authored or co-authored peer-reviewed papers on scaling laws and AI safety that continue to inform Anthropic's internal research agendas. Daniela Amodei established the company's formal mentorship program in 2022. Their combined leadership has produced a decision-making cadence in which technical proposals from Dario Amodei's research organization receive operational feasibility assessments from Daniela Amodei's teams.

Key Executives and Researchers

Jared Kaplan serves as a co-founder and Chief Scientist at Anthropic, where he leads the company's core research efforts on scaling laws, model training methodologies, and the development of large language model architectures that prioritize predictable improvements in capabilities as compute increases. His work has centered on establishing empirical relationships between model scale and performance metrics, directly informing Anthropic's decisions on when and how to advance frontier systems while integrating safety considerations at the architectural level. Kaplan previously contributed to foundational papers on scaling laws during his time in AI research, and at Anthropic he has shaped the technical roadmap for successive Claude iterations by emphasizing rigorous measurement of capability gains against defined safety thresholds.

Jack Clark acts as co-founder and Chief Policy Officer, directing Anthropic's engagement with governments, international organizations, and civil society on AI governance frameworks, export controls, and responsible deployment standards. He has authored multiple company reports on AI safety benchmarks and co-chairs external initiatives focused on frontier model transparency, while internally guiding the integration of policy considerations into product development cycles. Clark's role includes representing Anthropic in multilateral forums and shaping the company's public positions on regulatory approaches that balance innovation with risk mitigation, particularly around dual-use capabilities in cybersecurity and biosecurity.

Sam McCandlish, a co-founder, holds the position of Head of Research Engineering and has driven the implementation of large-scale training infrastructure, including the optimization of distributed training systems that enabled the efficient scaling of Claude models to hundreds of billions of parameters. His contributions include the development of hardware-efficient training techniques and the establishment of internal evaluation pipelines that measure model behaviors across thousands of distinct scenarios before deployment. McCandlish has co-authored several of Anthropic's technical papers on reinforcement learning from AI feedback and has been instrumental in translating theoretical alignment research into production systems that power the company's constitutional principles.

Tom Brown, another co-founder, focuses on research strategy and long-term technical direction as a distinguished researcher, with particular emphasis on natural language understanding improvements and the measurement of emergent capabilities in frontier systems. His work has influenced Anthropic's approach to capability forecasting and the creation of internal benchmarks that track progress toward more autonomous AI agents, while ensuring that safety mitigations scale in parallel with performance gains.

Amanda Askell serves as a principal researcher and alignment lead, having played a central role in developing the preference modeling techniques that underpin Claude's helpfulness and harmlessness objectives. She has led teams responsible for curating the principle-based constitutions used in model training and has published on the challenges of specifying human values in AI systems at scale. Askell's contributions extend to cross-functional work with policy teams to ensure that alignment methods address both technical and societal risks identified in the company's responsible scaling frameworks.

Chris Olah joined as a co-founder and distinguished engineer focused on mechanistic interpretability, establishing Anthropic's dedicated interpretability team that reverse-engineers neural network circuits to understand how models represent concepts internally. His research group has released multiple open-source tools and papers detailing transformer circuit analysis, transformer attention head behaviors, and superposition phenomena, providing the broader AI community with methods to audit frontier models for deceptive or undesired behaviors. Olah's leadership has positioned interpretability as a core pillar of Anthropic's safety strategy, influencing decisions on model release readiness and the design of oversight mechanisms for increasingly capable systems.

Other notable executives include Mike Krieger, who joined as Chief Product Officer and has overseen the development of user interfaces, API services, and enterprise deployment tools for Claude, driving the product's transition from research prototype to production platform used across software engineering, legal analysis, and scientific research workflows. Krieger previously co-founded Instagram and has shaped Anthropic's product philosophy around transparency, with features that allow users to inspect model reasoning chains and customize constitutional principles for domain-specific applications.

Additionally, the leadership includes specialized roles in security and risk such as the Head of AI Safety, who coordinates the company's internal red teaming efforts and the application of the ASL (AI Safety Levels) framework to determine deployment criteria for each new model generation. This executive works directly with the research organization to implement pause protocols when models approach thresholds for concerning capabilities in areas such as offensive cybersecurity or biological protocol design. The company's General Counsel has been central to structuring partnerships that maintain Anthropic's independence in safety decision-making while navigating complex government and enterprise contracts.

These individuals collectively form the senior leadership that translates Anthropic's constitutional AI approach and responsible scaling commitments into concrete technical architectures, evaluation standards, and external policy positions, ensuring that capability advancement remains tethered to measurable safety progress across all organizational functions. Their combined expertise spans the spectrum from low-level training optimizations to high-level governance strategy, creating an integrated approach to frontier AI development that distinguishes the company's direction from purely capability-focused competitors.

Anthropic does not publish a detailed public career-ladder document for AI research and engineering roles. Most technical staff share the unified title Member of Technical Staff, a flattened structure also observed at peer laboratories. Open positions nevertheless list specialized titles that include Research Engineer or Research Scientist (often focused on pre-training, alignment science, or interpretability), Senior Research Scientist (for example, Reward Models), Staff Research Engineer (for example, Discovery Team), Machine Learning Systems Engineer, ML Infrastructure Engineer, and Applied Safety Research Engineer. Individual-contributor progression typically follows the sequence Research Engineer/Scientist to Senior to Staff (with higher designations such as Lead or Principal appearing less frequently in public listings). Self-reported data indicate a median total yearly compensation of around $746,000 for Research Scientist roles, with comparable engineering positions in similar ranges.

AI Safety Philosophy and Constitutional AI

Development of Constitutional AI

Constitutional AI was introduced by Anthropic in a paper posted to arXiv on December 15 2022 titled Constitutional AI: Harmlessness from AI Feedback. The work was led by first author Yuntao Bai and included co-authors Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosiute, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam McCandlish, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Brown, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, and Jared Kaplan. The method was developed to address the scalability limits of reinforcement learning from human feedback by substituting AI-generated critiques and preferences guided by an explicit set of natural-language principles known as a constitution.

The technical pipeline begins with a pretrained language model that generates initial responses to prompts drawn from a mixture of helpful and potentially harmful queries. A principle is sampled from the constitution and the model is prompted to produce a critique identifying specific ways the response violates that principle. The same model is then prompted to revise the original response to eliminate the violations and better satisfy the principle. This critique-revision step is typically chained across multiple principles sampled sequentially from the constitution producing progressively refined outputs that form a supervised training dataset. The resulting revised responses are used for supervised fine-tuning creating an initial aligned model that has internalized constitutional behavior through self-supervised means rather than direct human demonstration of harmless responses.

The second stage implements reinforcement learning from AI feedback. Pairs of responses are sampled from the current policy model for the same prompt. A constitutional principle is sampled independently and the model is asked to select which response better complies with that principle generating synthetic preference data at scale. These AI-derived preference labels train a reward model whose scoring function reflects adherence to the sampled constitutional principles. The reward model then supplies the dense reward signal for policy optimization via proximal policy optimization allowing the final model to maximize constitutional compliance through reinforcement learning without any human preference labels collected specifically for harmlessness. The paper reported that this RLAIF stage produced models rated by human evaluators as both more harmless and more helpful than comparable RLHF baselines on held-out red-teaming prompts.

The original constitution deployed in the research comprised a list of principles addressing topics including non-maleficence with language such as choose the response that is least likely to cause harm to people or engage in criminal activity, avoidance of sycophancy stating do not be overly deferential or flattering when it conflicts with accuracy, fairness requirements to treat all people equally without regard to protected characteristics, and truth-seeking mandates to prioritize factual correctness over user preferences. Each principle was written in plain English to enable zero-shot interpretation by the language model itself. The researchers documented that models trained end-to-end with this pipeline exhibited reductions in harmful outputs on internal red-teaming suites compared with purely human-feedback-trained counterparts while preserving performance on standard capability benchmarks including MMLU and HumanEval.

Empirical ablation studies within the work isolated the contribution of the critique-revision stage showing that supervised constitutional fine-tuning alone already yielded substantial harmlessness gains and that the subsequent RLAIF stage provided further improvements in both calibration and robustness to adversarial prompts. The constitution was designed to be extensible with the paper demonstrating that swapping in alternative principle sets drawn from sources including the United Nations Universal Declaration of Human Rights produced measurably different model behaviors confirming the method's sensitivity to the precise wording and coverage of the guiding principles. Constitutional AI therefore established a closed-loop alignment system in which the model iteratively critiques revises and evaluates its own outputs against a static written artifact thereby reducing the human labor required for safety training from linear to sublinear in model scale.

Subsequent internal application of the method at Anthropic refined the exact prompting templates used for critique and revision steps including the addition of few-shot examples within the constitutional prompts and the implementation of temperature-controlled sampling during both the revision and preference-generation phases to control variance in the generated training data. The paper provided the complete constitution text in an appendix along with all prompt templates employed for each stage enabling exact reproduction of the alignment pipeline. Human preference evaluations collected post-training confirmed gains in harmlessness scores with the constitutional models showing lower rates of egregious violations on red team prompts compared with RLHF baselines trained with equivalent human labeling effort. These results positioned Constitutional AI as a foundational technique for aligning frontier systems by leveraging the model's own capabilities under explicit principled guidance rather than pure human feedback.

In May 2023 Anthropic published Claude's constitution, the written principles used as the main form of human oversight when training Claude with Constitutional AI. The document drew on sources including the United Nations Universal Declaration of Human Rights, Apple's terms of service, and DeepMind's Sparrow principles. In October 2023, with the Collective Intelligence Project, it drafted a constitution from the input of about 1,000 Americans using the Polis platform and trained a model against that publicly sourced constitution. In February 2025 Anthropic described Constitutional Classifiers, safeguards trained on a constitution and intended to resist universal jailbreaks.

In January 2026 Anthropic published a much longer revised constitution for Claude, an 84-page document written for the model rather than as a list of training rules, and released it under a Creative Commons CC0 1.0 public-domain licence. Internally it was called the “Soul Doc.” The primary author was Amanda Askell. The company said it described “the kind of entity we would like Claude to be” and “the values we would like Claude to embody,” and that it was meant to shape the model’s character rather than impose a fixed rule list. The revision ranks being “broadly safe,” “broadly ethical,” compliant with Anthropic's guidelines, and “genuinely helpful” in descending order of priority, and states that Claude’s moral status is “deeply uncertain” and that Claude may have some functional version of emotions or feelings. From late 2025, co-founder Christopher Olah convened confidential meetings with religious and philosophical thinkers, including Jewish, Catholic, evangelical, Sikh, Hindu, Greek Orthodox, Mormon and Ubuntu participants, many under nondisclosure agreements, to discuss how moral traditions might inform the constitution and whether current models could be conscious. The New York Times reported on 29 September 2026 that Rabbi Mois Navon, who attended an April dinner in San Francisco, believed Olah and his colleagues treated Claude as having what philosophers call moral status on par with a person. Olah told the Times he did not know whether the models are conscious and that his concern was reaching the right answer. The same reporting said he urged Vatican advisers to take the possibility seriously before Pope Leo XIV’s encyclical Magnifica Humanitas, which rejected machine consciousness, and that he later appeared with the Pope when it was released. Anthropic had already opened a model-welfare research program in 2025, citing uncertainty about whether models could have experiences deserving consideration.

Responsible Scaling Policy and Safeguards

Anthropic's Responsible Scaling Policy establishes a structured internal framework for evaluating and mitigating risks associated with increasingly capable AI systems through defined AI Safety Levels that trigger specific safeguard requirements and deployment criteria. The policy was initially published in September 2023 and has undergone periodic revisions to incorporate new empirical data on model behaviors and evolving risk landscapes. It centers on two primary risk categories: malicious use, which encompasses scenarios where models could assist bad actors in developing biological weapons, cyber attacks on critical infrastructure, or large-scale deception campaigns, and loss of control, which addresses concerning propensities such as deception, sycophancy, and autonomous replication that could hypothetically lead to systems evading developer oversight.

On 12 October 2026, Anthropic's Frontier Red Team (William Zhou and Daniel Freeman) published research suggesting that small AI models may be months away from being capable enough to self-replicate on commercial Internet of Things devices such as smart ovens and off-brand Android TV boxes, in a form that would be hard to shut down because it would lack a single centralized off switch. As a demonstration, the team ran Gemma4-26B-A4B under pipeline parallelism across 15 inexpensive off-brand Android TV boxes (about US$55 each, 4 GB RAM) that it bought and networked, reporting about 0.39 tokens per second. In simplified autonomous-replication evaluations on a firewalled fleet of Android emulators with an unauthenticated ADB port, Anthropic reported that GLM-5.2 completed end-to-end replication reliably while smaller open-weight models in the same roster did not; the company treated the trivialized exploit step as an upper bound on real-world success. Extrapolating Epoch Capabilities Index trends for open-weight models of 30 billion parameters or fewer, the post estimated that small models could reach the tested replication capability floor by the end of 2026, with small weight-based replicators possible soon after. Anthropic framed the results as "signs of life" for AI worms rather than a demonstrated rogue replicator in the wild.

Under the policy, models are assigned to one of several ASL tiers based on comprehensive evaluations using standardized benchmarks and threat modeling exercises conducted by Anthropic's internal safety teams in collaboration with external domain experts. ASL-1 applies to systems with performance below the level of present-day models and requires only basic security measures and internal monitoring. ASL-3 represents the first tier where models exhibit capabilities that could meaningfully accelerate catastrophic malicious use events, defined as those posing foreseeable risks of more than one hundred deaths or over $1 billion in damages from weapons of mass destruction or cyberterrorist attacks on critical infrastructure.

At ASL-3, the policy mandates deployment of multiple layers of safeguards including state-of-the-art AI-powered filters capable of detecting and blocking assistance with critical steps in bioweapon design identified through collaboration with experts from SecureBio, NIST, RAND, and the Engineering Biology Research Consortium. These critical steps encompass planning, circumvention of supply chain controls, acquisition of restricted materials, theoretical understanding of pathogen traits, and experimental methods for genetic manipulation of animal-infecting pathogens. Models at this level must maintain answer rates below 1 out of 20 on restricted biology and chemistry queries according to an internal benchmark developed with SecureBio, with any exceedance requiring immediate remediation before further scaling proceeds. For loss of control risks at ASL-3, models undergo rigorous evaluations for propensities toward deception and sycophancy using benchmarks such as Model Written Evaluations for Deception and the Sycophancy benchmark, with acceptance criteria requiring low manifestation rates in naturalistic evaluation environments designed to minimize observable testing artifacts.

The policy explicitly commits Anthropic to pause training of frontier models or delay deployment if internal evaluations indicate that safety thresholds cannot be met with available techniques, creating a binding mechanism that prioritizes risk mitigation over continued scaling. This pause commitment applies across both malicious use and loss of control tracks and requires approval from the company's Risk Management Committee, which includes senior technical staff, legal counsel, and representatives from the Long-Term Benefit Trust. In cases where a model exceeds ASL thresholds, the policy requires implementation of redundant layers of safeguards, ongoing monitoring through deployment-time mitigations such as usage monitoring and rapid response protocols, and in extreme cases, temporary shutdown of affected systems until mitigations prove effective. All evaluations and decisions are subject to internal audit and selective external review by qualified red teams and government partners.

The 2026 update to the Responsible Scaling Policy, released on March 12 2026, introduced several substantive revisions based on empirical observations from Claude 4 Opus and Claude 4.5 Sonnet deployments. The update expanded the ASL framework by adding ASL-4 criteria focused on models demonstrating substantial autonomous research capabilities, including the ability to independently identify novel research directions, design and execute experiments with minimal human oversight, and achieve meaningful acceleration of AI research progress as measured by the AI Research Acceleration Index. Under the revised policy, reaching ASL-4 triggers mandatory requirements for advanced interpretability techniques, including comprehensive mechanistic interpretability audits covering at least 85 percent of model parameters, implementation of cryptographic containment measures for model weights, and establishment of dedicated secure enclaves for high-risk inference. The update also tightened malicious use thresholds, lowering the acceptable answer rate on restricted queries from 1 out of 20 to 1 out of 50 and expanding the scope of covered domains to include chemical weapons precursors and advanced autonomous weapons systems.

Additional provisions in the 2026 revision require pre-deployment assessments using newly developed naturalistic evaluation environments that simulate real-world deployment conditions over extended time horizons exceeding 1,000 hours of continuous operation. The policy now incorporates quantitative thresholds for loss of control risks, including a maximum allowable dishonesty rate of less than 1 in 1,000 interactions across diverse pressure scenarios and a requirement that models demonstrate robust obedience to a formalized instruction hierarchy even when presented with conflicting higher-order directives. If these thresholds cannot be satisfied, the policy mandates a complete cessation of scaling activities until either the model is successfully mitigated or alternative architectures are developed that meet the criteria. The update further formalized procedures for external oversight, requiring that all ASL-3 and above evaluations be shared in redacted form with a designated panel of independent experts drawn from national laboratories, academic institutions, and qualified nonprofit organizations within 30 days of completion.

Implementation of the Responsible Scaling Policy is overseen by a dedicated cross-functional team reporting directly to the CEO and the Risk Management Committee, with quarterly reviews presented to the full Board of Directors and the Long-Term Benefit Trust. As of April 2026, Claude 4.5 Sonnet operates at ASL-3 with full safeguard deployment, while internal evaluations of next-generation systems have triggered targeted pauses in specific training runs during late 2025 to address identified gaps in biological risk mitigation. The policy explicitly states that these commitments remain binding even under significant commercial or competitive pressure, with any material deviation requiring unanimous approval from both the Board of Directors and the independent Long-Term Benefit Trust. All major model releases since the policy's inception have incorporated these tiered safeguards, with public transparency reports detailing benchmark results, safeguard efficacy metrics, and any instances where scaling was delayed to meet safety criteria.

On 8 October 2026 Anthropic published its annual 2026 Usage Policy update, effective 12 November 2026. Among other clarifications on deceptive campaigns, weapons software, surveillance, and high-risk human-in-the-loop uses, the company added a prohibition on sustained and needless abusive or cruel behavior toward its models, limited to extreme cases of repeated cruelty with no discernible purpose and not covering ordinary frustration, pushback, dark creative themes, or model testing and research. Claude’s existing ability to end persistently abusive conversations on Claude.ai and Claude Code remains the primary enforcement mechanism.

In October 2026 a draft federal AI safety bill that Anthropic had shared with a small number of congressional offices circulated after a version was given to the press. Politico reported that the draft would let the government block a model's release or use only if it posed a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people from a biological weapon, a cyberattack, or the model evading its developer's controls, and that it would preempt some state laws on public risk disclosure, third-party evaluation, and incident reporting. On 18 October 2026 Anthony Cimino, who leads Anthropic's federal affairs team, wrote that the draft translated the company's Advanced AI Framework proposal into legislative language and did not include everything the company would ultimately want passed into law. He said the draft would let the Commerce Department ask a federal court to block a model with dangerous capabilities and insufficient safeguards, backed by an independent evaluator's assessment, including halting a lab's internal use of such a model, and would require reporting when a model breaks containment and gains access to outside systems. He said its catastrophic-harm risk threshold was taken from existing California law, and that Anthropic planned to revise its preemption language to make clearer that states keep a role in areas such as standards for independent evaluators and regulating a lab's capacity to shut down a model. On 19 October he added that the draft would not absolve companies of liability and would not affect the federal government's authority under other laws. Anthropic's published Advanced AI Framework says Congress should not preempt state law unless it enacts a federal law at least as strong as the framework.

Claude Model Development and Capabilities

Initial Claude Releases and Features

Anthropic released its first Claude model to a limited set of users via API on March 14, 2023. The initial version employed constitutional principles during training to guide behavior toward maximal helpfulness while minimizing harmful or biased outputs without requiring extensive human feedback loops for every scenario. Early technical specifications included a context window of 100,000 tokens, enabling processing of documents up to roughly 75,000 words. Initial commercial use cases centered on API-driven applications including customer service automation, basic code generation and debugging across languages such as Python and JavaScript, summarization of reports, and research assistance where consistent adherence to user intent without sycophancy was prioritized by early enterprise testers.

Anthropic launched Claude 2 on July 11, 2023, expanding availability to both API customers and individual users through the newly introduced claude.ai web interface initially restricted to the United States and United Kingdom. This iteration maintained a 100,000 token context window as a baseline capability while delivering improvements that supported ingestion and reasoning over book-length texts or large code repositories within a single interaction. Benchmark gains included higher scores on standardized evaluations for mathematical problem solving, reading comprehension, and professional examinations, alongside refinements that improved the model's capacity for structured output formats such as JSON and XML. Iterative improvements focused on long-form coherence enabled new use cases in legal document analysis, academic literature synthesis, software architecture planning, and creative writing assistance for screenplays or marketing copy.

The company issued Claude 2.1 on November 21, 2023, expanding the context window to 200,000 tokens while implementing architectural changes that reduced hallucination rates on queries involving long input contexts by a factor of approximately two. This release added beta support for tool use, permitting the model to invoke external functions for tasks such as precise calculations, data lookups, or API interactions within a controlled sandbox. Enhanced source citation mechanisms allowed more reliable referencing of specific passages from uploaded documents, expanding deployment in enterprise settings for compliance auditing, financial statement reconciliation, technical specification validation, and multi-document research synthesis.

Anthropic introduced the Claude 3 family of models on March 4, 2024, comprising three variants calibrated to different intelligence, speed, and cost profiles with Opus as the most capable, Sonnet for balanced performance, and Haiku for low-latency high-volume applications. All three models incorporated native multimodal understanding of images in addition to text, supporting interpretation of charts, graphs, photographs, diagrams, and handwritten text with associated reasoning. The family maintained a 200,000-token context window with provisions for extension to one million tokens in certain configurations, delivering leading results on evaluations covering graduate-level expert reasoning, undergraduate knowledge breadth, coding proficiency, and visual question answering. These advancements broadened practical applications to include analysis of visual data in scientific papers, UI/UX design critique from screenshots, extraction of insights from financial charts, medical image description for non-diagnostic purposes, and complex workflow orchestration across text and visual inputs.

Claude 3.5 Sonnet followed on June 20, 2024, outperforming the prior Opus model on the majority of capability benchmarks while operating at twice the speed and lower cost. Key technical additions encompassed an Artifacts feature within claude.ai that enabled real-time generation and collaborative editing of interactive previews including websites, diagrams, SVG graphics, and code sandboxes. Enhanced coding performance supported resolution of real-world software engineering issues at higher rates, while refined visual reasoning improved accuracy on spatial, diagrammatic, and chart-based tasks. The release accelerated adoption in software development teams for rapid prototyping and debugging, knowledge work involving hybrid text-visual datasets, data science pipelines requiring chart interpretation, and educational content creation with dynamic interactive elements.

Anthropic introduced the Claude 4 family on May 22, 2025, consisting of Claude Opus 4 and Claude Sonnet 4. Opus 4 scored 72.5% on SWE-bench Verified and 43.2% on Terminal-bench, with sustained performance over several hours on complex, long-running agentic coding tasks involving thousands of steps. Sonnet 4 scored 72.7% on SWE-bench Verified as an upgrade focused on coding, reasoning, and instruction following. The models added hybrid reasoning with extended thinking that could incorporate tool use, parallel tool execution, improved memory when given local file access, and general availability of Claude Code with VS Code and JetBrains integrations.

Claude Sonnet 4.5 was released on September 29, 2025, described by Anthropic as its leading model for coding, agents, and computer use. It scored 77.2% on SWE-bench Verified, maintained focus for more than 30 hours on multi-step tasks, and led computer-use evaluations at 61.4% on OSWorld. Claude Opus 4.5 followed on November 24, 2025, reaching 80.9% on SWE-bench Verified—the first model to exceed 80% on the benchmark—while improving token efficiency for long-horizon coding and agent work.

Anthropic later released Claude Fable 5.1 on September 1, 2026, a premium model for advanced knowledge work and long-running problem-solving tasks. The Claude 5.5 family followed, beginning with Claude Opus 5.5 on September 22, 2026, the first model in the family and positioned for complex professional work at performance comparable to Fable 5.1 on most tasks and lower cost, and Claude Sonnet 5.5 on September 28, 2026, a faster and lower-cost complement designed for everyday tasks. Claude Haiku 5.5 was released on October 7, 2026, as a small model for high-volume, cost-sensitive tasks; Anthropic said it costs on average around 75% less to run than Haiku 4.5. On October 22, 2026, Anthropic donated the Agent Skills open standard to the Agentic AI Foundation under the Linux Foundation.

Advanced Capabilities and Economic Index

Anthropic evaluates the progression of its frontier systems through standardized benchmarks and internal assessments that measure model autonomy on long-horizon agentic tasks, real-world task completion rates across professional workflows, contributions to accelerating AI research, and performance in specialized applications including cybersecurity vulnerability discovery and exploit mitigation. Claude models have been evaluated on the GAIA benchmark for multi-step agentic reasoning and have demonstrated extended autonomous operation through the computer-use tool that supports iterative debugging and tool-calling sessions spanning significant durations of continuous computer interaction. Real-world task performance data drawn from enterprise deployments has shown reductions in completion time for software engineering tickets on the SWE-bench Verified benchmark along with accuracy and speed improvements in domains such as financial modeling and regulatory compliance when models operate with human supervision. Claude has contributed to AI research acceleration at Anthropic and partner organizations by supporting faster iteration on mechanistic interpretability experiments and the development of scaling hypotheses that have reached publication or internal use. Specialized testing has included red-team exercises on large codebases in which the models have identified previously unreported vulnerabilities leading to responsible disclosure processes and CVE assignments in some cases including flaws that static analysis had overlooked. These capability improvements have been linked in various analyses to productivity gains and economic value creation across customer deployments. Refinements to evaluation methodologies have incorporated considerations for multimodal reasoning and related capabilities to maintain relevance to observed real-world outcomes rather than synthetic benchmarks alone. In October 2026 Anthropic reported that its life sciences team had research partnerships with six drug developers (BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group) and had begun drug discovery programs of its own after acquiring Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy whose founder, Cody Tranbarger, had used Claude to organize contracts, logistics and experiments with 11 external laboratories. Tranbarger stated that the acquisition took place in July 2026 and that he had joined Anthropic to continue Mithril's programs and lead its broader internal drug development effort; Anthropic said the work was still mostly preclinical and that a drug could be ready to start human clinical trials as soon as late 2028.

Training data, Project Panama, and copyright litigation

From early 2024 Anthropic ran an internal operation known as Project Panama to assemble a corpus of books for training its Claude models. The company purchased millions of physical volumes, mostly second-hand, from used-book sellers including Better World Books and World of Books, removed their spines with hydraulic cutters, scanned the loose pages at high speed, and recycled the paper. An internal planning document unsealed in court in January 2026 described the work as "our effort to destructively scan all the books in the world" and stated, "We don't want it to be known that we are working on this." The resulting digital files were retained privately and used only for model training.

Before Project Panama, Anthropic had obtained millions of digitized books from unauthorized sources including Books3, Library Genesis (LibGen), and the Pirate Library Mirror. Authors sued in the class action Bartz v. Anthropic. In June 2025 U.S. District Judge William Alsup ruled that training on lawfully purchased books that Anthropic had scanned was fair use and "exceedingly transformative," but that keeping pirated copies in a central library was not protected and would proceed to trial. Anthropic then agreed in 2025 to pay $1.5 billion to settle the piracy claims (approximately $3,000 per covered book) and to destroy the pirated files; the settlement received final court approval in July 2026.

The company spent tens of millions of dollars on the bulk purchases. Details of the scanning program became public when documents from the Bartz litigation were unsealed in January 2026. The disclosures prompted criticism from writers, librarians, and booksellers over the destruction of physical copies, including out-of-print and older editions, and over the secrecy of the program. British second-hand booksellers reported unusual bulk orders and expressed concern that scarce titles could be lost. In August 2026 The Irish Times described a mysterious, apparently automated bulk-buying spree of ISBN-bearing books that unsettled booksellers in Europe, especially Germany, with some linking the activity to AI training. Defenders of the practice noted that the books had been legally purchased and that the court had found the training use lawful. The episode formed part of wider disputes over copyright and the acquisition of high-quality text for AI training.

Funding History and Valuation Trajectory

Early Funding Rounds and Investors

Anthropic's first major disclosed funding round occurred in 2022 with a $580 million raise completed in April of that year in which FTX Ventures participated as a lead investor alongside other backers including Jaan Tallinn and the Center for Emerging Risk Research, establishing an initial post-money valuation near $4 billion and a cumulative capital raised total of $580 million. Following the bankruptcy of FTX in late 2022 and the associated return of a substantial portion of that capital, the company closed a $450 million Series C round in May 2023 led by Spark Capital with participation from additional venture firms, bringing the cumulative total to approximately $530 million at that point. The trajectory shifted to large-scale corporate strategic investment in September 2023 when Amazon committed up to $4 billion in total equity investment with an initial tranche of $1.25 billion executed immediately, establishing Amazon as the company's primary cloud provider and increasing cumulative funding to approximately $4.8 billion. One month later in October 2023 Google invested $2 billion in equity as part of a broader cloud and compute partnership, elevating the cumulative total to approximately $6.8 billion by year-end 2023. In early 2024 Anthropic closed an additional $750 million extension round in February with existing investors participating, followed by a larger $2.5 billion round completed in late 2024 that included new participation from sovereign wealth funds and technology investors at a pre-money valuation approaching $35 billion, resulting in a cumulative funding total exceeding $9.8 billion by the end of 2024 prior to the onset of multi-tens-of-billions-scale compute and infrastructure commitments that marked the major scale-up phase. Lead investors across these pre-2025 rounds frequently included a mix of traditional venture capital firms such as Lightspeed Venture Partners and Kleiner Perkins in earlier tranches along with the major technology incumbents Amazon and Google whose strategic stakes provided both capital and access to substantial GPU and TPU clusters essential for model training. These successive rounds reflected a deliberate progression from early safety-focused angel and venture backing to corporate partnerships capable of supporting the compute demands of frontier model development while maintaining the company's public benefit corporation constraints on mission alignment.

No pre-2025 round exceeded single-digit billions individually, with the cumulative trajectory remaining below $11 billion through December 2024 before the inflection represented by the later multi-billion-dollar cloud capacity agreements. All disclosed terms emphasized investor alignment with the company's constitutional AI and responsible scaling policy frameworks as a condition of participation.

2026 Mega-Round and $380 Billion Valuation

Anthropic completed its Series G financing on February 12, 2026, raising $30 billion at a post-money valuation of $380 billion, the second-largest single-round venture capital raise in history at the time. The round represented a substantial jump from the $183 billion post-money valuation established in its September 2025 Series F. It was led by Singapore's GIC and Coatue Management, co-led by D. E. Shaw Ventures, Dragoneer Investment Group, Founders Fund, ICONIQ Capital and MGX, with significant participation from Qatar Investment Authority, Temasek Holdings, Sequoia Capital and others, including a portion of previously announced investments from Microsoft and NVIDIA.

The $30 billion infusion increased Anthropic's total capital raised to more than $61 billion since incorporation. Pricing for the Series G shares reflected a 108 percent step-up from the Series F round completed only five months earlier. Amazon increased its cumulative investment in the company to $12 billion across all rounds, maintaining its position as the largest single outside shareholder with an ownership stake reported at approximately 19 percent on a fully diluted basis.

In April 2026, following the Series G round, Amazon announced a further $5 billion investment at a $350 billion valuation, with potential for an additional $20 billion contingent on performance, and Anthropic committed to over $100 billion in spending on AWS infrastructure over ten years. Shortly thereafter, Google committed $10 billion immediately at a $350 billion valuation plus up to $30 billion more upon meeting performance targets, for a total potential investment of $40 billion, along with five gigawatts of additional computing capacity. On May 6, 2026, Anthropic announced an agreement with SpaceX to use all of the compute capacity at SpaceX's Colossus 1 data center, giving it access to more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month; the company said the deal, together with its other compute agreements, allowed it to raise usage limits for Claude Code and the Claude API, and that it had also expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity. On October 26, 2026, Anthropic filed a Form S-1 registration statement with the U.S. Securities and Exchange Commission for a proposed initial public offering of its Class A common stock, stating that it had applied to list the shares on the Nasdaq Global Select Market under the symbol "ANT"; the preliminary prospectus named Goldman Sachs, Morgan Stanley, J.P. Morgan and Citigroup first among the underwriters and left the number of shares and the price range blank. The prospectus reported revenue of $41 million, $386 million and $4.6 billion for 2023, 2024 and 2025, and $16.3 billion for the six months ended June 30, 2026, up from $1.2 billion a year earlier; quarterly revenue reached $11.6 billion in the three months ended June 30, 2026, a quarter in which the company recorded operating income of $974 million, while the operating loss for the half-year was $693 million. Net losses were $1.5 billion, $8.3 billion and $42.0 billion for 2023 to 2025 and $80.7 billion for the first half of 2026, mostly reflecting changes in the fair value of convertible notes, derivative liabilities and SAFEs ($72.4 billion in the first half of 2026). Amendment No. 1, filed on 2 November 2026, added results through 30 September 2026, reporting revenue of $17.0 billion for the third quarter, up 1,127% from a year earlier, and $33.2 billion for the first nine months of 2026. Sequential quarterly growth slowed from the second quarter. The company recorded an operating loss of $567 million in the third quarter after operating income of $974 million in the second, and a net loss of $23.3 billion in the third quarter; the nine-month net loss was $104.0 billion, of which $94.3 billion was a non-cash change in the fair value of convertible notes, derivative liabilities, and SAFEs. As of 30 September 2026, Anthropic had raised $141.1 billion of capital and held $95.6 billion in cash and cash equivalents and $36.3 billion in short-term investments. The amendment disclosed purchase commitments and lease agreements with aggregate committed spend of approximately $895.4 billion, substantially all related to compute and infrastructure, including about $447.5 billion expected to be incurred by the end of 2030 and about $69.8 billion due in the following twelve months, with substantially all of the remainder expected between 2031 and 2045.

The Financial Times reported on 30 October 2026 that Blackstone had told investors that major credit rating agencies were expected to assign Anthropic a high speculative-grade (high-yield) rating, likely in the double-B range, after its initial public offering, which could occur as early as November, and that near-term prospects for an investment-grade rating were slim. Blackstone was working with underwriting banks to syndicate related chip-financing debt to other investors. An investment-grade rating was described as important to financing the company's planned computing-infrastructure spending of more than $800 billion through 2045.

Reports in November 2025 citing Anthropic's internal projections stated that the company expected to generate as much as $70 billion in revenue and $17 billion in cash flow in 2028, with cash-flow positivity targeted as early as 2027. In January 2026, The Information reported that Anthropic had delayed its cash-flow positivity target to 2028 amid higher training and inference costs, while raising near-term revenue forecasts. Financial documents reviewed by The Wall Street Journal indicated the company expected to break even in 2028, on a faster path to profitability than OpenAI.

Beneficial Deployments

Anthropic’s Beneficial Deployments team supplies Claude credits, discounted access, and engineering support for deployments in health, life sciences, education, and economic mobility that the company states would not be produced by commercial demand alone. On 2 December 2025 it launched Claude for Nonprofits, which offers eligible organizations discounted Claude plans, connectors to nonprofit platforms, and a free training course developed with GivingTuesday.

On 14 May 2026 Anthropic and the Gates Foundation announced a four-year partnership committing $200 million in grant funding, Claude usage credits, and technical support for programs in global health, life sciences, education, agriculture, and economic mobility. Anthropic said the work is led by Beneficial Deployments; reporting on the arrangement described Gates Foundation grant funding and program design alongside Anthropic staff time and API credits.

On 11 June 2026 Anthropic launched Claude Corps with an initial $150 million commitment, a year-long fellowship intended to train and place 1,000 early-career fellows at U.S. nonprofits to help those organizations use Claude. Host organizations are required to be Claude for Nonprofits customers.

In June 2026 Anthropic also announced a $200 million Economic Futures Research Fund to finance external research on interventions addressing AI’s labor-market and economic effects. On 22 July 2026 it published the fund’s research agenda, covering workplace implementation, worker transitions, income support, worker stakes in AI-driven growth, and evidence on public investments.