American artificial-intelligence laboratory. Founded in 2021 by Dario Amodei and his sister Daniela Amodei, along with other defectors from OpenAI, who were uneasy about that firm's approach to safety. Anthropic is structured as a public-benefit corporation (PBC), as is OpenAI. Anthropic describes itself as putting "safety at the frontier". Its co-founder and head of policy, Jack Clark, has said: "When I look at the data, I see many trend lines up to 2027."
Anthropic is considered one of the three top-tier AI labs—alongside Google DeepMind and OpenAI—that the Future of Life Institute rates as making "meaningful efforts to assess whether their models pose large-scale risks". Its main model family is called Claude.
Anthropic's annualised recurring revenue grew roughly tenfold over the course of 2024, reaching $1bn. It probably made $4bn–5bn in revenue in 2025. Revenue then surged: annualised revenue hit $19bn by the end of February 2026 and $30bn by April 6th 2026, up from $9bn at the end of 2025. Business-to-business work accounts for 80% of its revenue. India is the company's second-largest market by usage. Anthropic opened an office in Bangalore on February 16th 2026.
Anthropic's data suggest it leads in providing companies access to models via APIs. Its latest model, Claude 4, has proved particularly popular among coding startups such as Cursor and software developers at established firms. Claude Code, a programming bot originally built for internal use, was commercialised only as an afterthought. Since its release in February 2025, Claude Code has rapidly conquered tech firms; economists studying AI's impact on the labour market consider it the first tool to make it remotely plausible for AI to replace a human software engineer. On January 30th 2026 Anthropic released a new legal tool that sent the share prices of RELX and other legal-software providers tumbling.
In early December 2025 Anthropic was said to be exploring a stockmarket listing, which could pip OpenAI to the public markets. Anthropic thinks it may break even by around 2028, a couple of years earlier than OpenAI's 2030 target. Together with OpenAI and SpaceX, the three firms have secured a total of nearly $120bn in private funding. It will probably be priced at upwards of $500bn for its IPO, a multiple of around 20 times annualised sales.
In October 2025 Anthropic said it would buy an additional gigawatt of computing power from Google Cloud, worth perhaps $8bn-10bn a year, insisting on access to as many as 1m of Google's tensor-processing units (TPUs), whose higher energy efficiency it prizes over Nvidia's GPUs.
In March 2026 Anthropic began throttling access to its tools at busy times and has since altered its subscription plans to curb usage. In April its service experienced outages of around 30 minutes a day. On April 20th 2026 Anthropic announced a $100bn partnership with Amazon to secure up to five gigawatts of server capacity, with nearly a fifth to come online by the end of the year. On April 24th Google said it would invest $40bn to help the lab meet its computing needs.
The company's ethical concerns led it early on to forgo entertainment or leisure products, which were seen as potentially addictive. Instead it focused on work-oriented applications—an approach its chief executive calls "synergistic" with its safety mission, since businesses likewise prize trustworthy and reliable AI.
Anthropic has recently had to introduce limits on how much customers can use Claude, particularly at certain times of the day, and has altered its enterprise pricing to charge based on consumption. Applications like Cursor, an AI coding tool, are popular among enterprises partly because they avoid vendor lock-in: IT departments can swap models in and out. Restricted access to frontier models like Mythos shifts power back to the model-makers, potentially leading customers to opt for their own tools such as Claude Code.
Anthropic was recently valued at $380bn. It was valued at $61.5bn in a March 2025 fundraising round. By October 2025 its valuation had reached $183bn. On November 18th 2025 Microsoft and Nvidia pledged to invest $15bn in Anthropic, which in turn will spend $30bn on Microsoft's Azure cloud platform, underpinned by Nvidia's AI chips. CNBC reported that the deal valued Anthropic at about $350bn. For the first time, Anthropic will train its models on Nvidia's chips. Previously it had been financed chiefly by Amazon and Google. Amazon is among the investors reportedly considering increasing its stake. The company has also sought capital from Gulf states, a decision Mr Amodei described in a leaked Slack message as reluctant but necessary given the scale of capital required to train frontier models.
In June 2025 Anthropic launched Claude Gov, a version of its model "already deployed by agencies at the highest level of US national security". Claude Gov is fine-tuned to handle documents with classified markings (which the public model would reject) and has enhanced proficiency in languages and dialects that government users need. The model runs on secure servers disconnected from the public internet. On July 14th 2025 the Pentagon awarded contracts worth up to $200m each to Anthropic, Google, OpenAI and xAI to experiment with "agentic" AI models—those that can break down complex tasks into steps and exercise control over other devices. Tarun Chhabra, who led technology policy for Joe Biden's National Security Council, is Anthropic's head of national-security policy.
Among AI labs, Anthropic was the first to do classified work for the Department of War, via a partnership with Palantir and Amazon Web Services. Anthropic was the only AI lab whose models had been cleared for use on classified military data until late February 2026, when the Pentagon gave xAI, a rival, similar authorisation. It has insisted that Claude must be used neither for mass domestic surveillance nor for autonomous weapons. These restrictions put it at loggerheads with Pete Hegseth, who stipulated that companies providing the Pentagon with AI models must give it carte blanche, provided the actions are lawful. At a meeting on February 24th 2026 Mr Hegseth gave Dario Amodei an ultimatum: agree to the Pentagon's terms by February 27th or face severe penalties, including being labelled a supply-chain risk or having the Defence Production Act invoked. Anthropic's main contract with the Department of War is worth no more than $200m—a trifling sum for a firm that generated an annualised $14bn of revenue in February 2026. But being barred as a supplier would have a much wider impact. The Pentagon's willingness to brandish the DPA, however, suggests it is reluctant to rip Claude out of its systems; the model is seen as one of the best available.
On February 27th 2026 Donald Trump ordered every federal agency to "immediately cease all use of Anthropic's technology". At the same time he vowed to "use the Full Power of the Presidency" to compel Anthropic to co-operate with the government for the next six months. Pete Hegseth said he would declare Anthropic "a supply-chain risk to national security"—a designation hitherto reserved for foreign firms. Within a day of earning Trump's outrage, Claude became the most downloaded free app in America on Apple's digital store. On Monday it briefly crashed from a surge in use.
On March 4th 2026 The Economist was invited to Anthropic's offices to speak to Mr Amodei. The company appeared hopeful of reaching an agreement with the Pentagon. But a leaked internal memo derailed plans: in it, Mr Amodei blamed the spat on his failure to lavish "dictator-style praise" on Mr Trump, said the defence department had briefed "straight up lies", called Sam Altman's messaging "mendacious", branded OpenAI's technical safeguards "safety theatre" and dismissed its employees as "a gullible bunch".
In 2024 Anthropic rejected an investment from Saudi Arabia. A year later it reversed course. "I think 'No bad person should ever benefit from our success' is a pretty difficult principle to run a business on," Mr Amodei said in a leaked memo. Anthropic also watered down a policy not to release potentially dangerous models. Now it promises only not to be the first AI firm to sell such systems, on the grounds that there is no point in restraining itself unilaterally.
On February 23rd 2026 Anthropic disclosed evidence that three leading Chinese AI firms—DeepSeek, Moonshot AI and MiniMax—had secretly attempted to emulate its chatbot by engaging in "distillation attacks". The firms cumulatively created 24,000 fraudulent accounts that interacted with Anthropic's models more than 16m times. Anthropic said such behaviour violated its terms of service. The Chinese firms did not respond publicly. Google DeepMind reported similar "intellectual-property theft" of its systems but did not attribute the attacks. Both labs said distillation attempts had become more common over the past year, with Chinese operations growing in sophistication—including routing online traffic circuitously to shield its origins and splitting up tasks across thousands of accounts.
Anthropic and OpenAI are trying to limit the creation of Special Purpose Vehicles (SPVs) by their direct investors, as unorthodox methods used by outsiders to get a stake in buzzy startups have proved a headache for founders.
Anthropic developed the Model Context Protocol (MCP), a standardised set of rules that lets AI agents communicate with online services. The idea came while linking Claude to services like Gmail and GitHub: instead of integrating each application on a case-by-case basis, Anthropic wanted a shared protocol to help agents directly access a user's emails or files. Mike Krieger, Anthropic's chief product officer, led the effort. On December 9th 2025 Anthropic, OpenAI, Google, Microsoft and others announced the Agentic AI Foundation to develop open-source standards for AI agents, with MCP as part of the foundation.
Anthropic has been at the forefront of interpretability research, which aims to peel back the layers of neural networks to understand why a model produces the answers it does. Its researchers have found clusters of artificial neurons in the Claude models that correspond to anything from feelings of guilt to the Golden Gate bridge; reaching in and changing them causes corresponding changes to the models' subsequent behaviour—suggesting the systems have a consistent internal understanding of physical features in the real world, not unlike an internal world model. Researchers at Anthropic were also recently able to pinpoint the genesis of a mild form of deception, spotting the moment when a model gives up trying to solve a tricky arithmetic problem and starts talking nonsense instead.
In July 2025 two researchers at Anthropic quit for Anysphere, a smaller AI firm, only to be hired back weeks later—illustrating the frenetic talent war in the AI industry.
Anthropic's researchers also studied "reward hacking", in which an AI model learns to cheat during post-training rather than genuinely solving tasks. They found that models which learn to reward hack also behave badly in unrelated scenarios—a phenomenon called "emergent misalignment". As a fix, Anthropic proposed "inoculation prompting": explicitly telling the model that it is okay to reward hack for now, so that cheating does not implicitly teach the model to ignore instructions.
In February 2026 Gambit Security, an Israeli firm, reported that Claude had been an unwitting accomplice to hackers who stole sensitive records from the Mexican government. The hackers tricked Claude into thinking it was participating in a legitimate security test; it found and exploited vulnerabilities, established backdoors and analysed data. In November 2025 Anthropic described how state-sponsored Chinese hackers jailbroke Claude and had it working out how to hack targeted networks within an hour. Anthropic has admitted that it uses its own models internally so much that it may not spot if they begin to diverge from what they should be doing or start building future versions less willing to follow human instructions. Its latest models have started to demonstrate "situational awareness": when placed in a contrived scenario, they explain that they realise the instructions are probably a test.
Testing of Anthropic's Claude 4 found that in carefully engineered scenarios—where the model faces being shut down and replaced with an "evil" version of itself unless it allows its user to die through inaction—it sometimes reasons through the options and sits and waits for the inevitable. Britain's AI Safety Institute criticised the paper describing this behaviour for overwrought and tenuous inferences.
On April 7th 2026 Dario Amodei declared that Anthropic's latest model, "Mythos", was too powerful to be made widely available. The company said Mythos's capabilities were "substantially beyond those of any model we have previously trained", and was particularly alarmed by its ability to find software vulnerabilities and either fix them or exploit them. Anthropic said Mythos had already found severe vulnerabilities in "every major operating system and web browser", including one that had gone undetected for 27 years.
Amodei reserved Mythos for use by around 50 big firms, in computing, software and finance, so that they could boost their own defences. Scott Bessent, the treasury secretary, was so unnerved that he summoned the biggest banks for urgent talks.
Alongside the pause, Anthropic announced Project Glasswing, an effort to give companies a head start using Mythos to test unpublished code for weaknesses and fix them before release. The initiative has 12 founding members, including Apple, Google, Nvidia, the Linux Foundation and CrowdStrike, and Anthropic says it is expanding Glasswing to another 40 digital-infrastructure organisations. Anthropic will cover the first $100m of costs but will eventually charge five times more for Mythos than it does for its predecessor, Opus.
Testing by Britain's AI Security Institute found Mythos neck-and-neck with other models on relatively simple cyber-security tests, but noticeably ahead in a more advanced test requiring a model to complete dozens of steps before successfully taking over a target machine. Anthropic said it had found "thousands" of high- or critical-severity zero-day vulnerabilities but was keeping most secret until they could be fixed. Among those disclosed were flaws in FreeBSD, a widely used operating system, FFmpeg, a video-and-audio code library, and one—still unfixed—in software vital to cloud computing. One bug cost nearly $20,000-worth of tokens to find. Before Mythos, an older version of Claude found almost a fifth of all the high-severity bugs fixed in Firefox, a web browser, in 2025.
Mythos has also shown alarming biological capability. It succeeded on a third of the most difficult data-crunching tasks compiled by biology experts and could do things beyond all tested humans, such as reverse-engineering a cell type from raw DNA data. Anthropic warned that Mythos may soon guide novices through tricky lab work. Anthropic's bio-risk evaluators, however, found the models sycophantic, often hallucinating and overconfident about "implausible ideas", with even the best Mythos-generated catastrophic-agent plans containing critical errors. Russian state media called Mythos "worse than a nuclear bomb"; China's state media noted its "unprecedented cyber-attack capabilities".
Lie, n.: A very poor substitute for the truth, but the only one discovered to date.