What Anthropic is
Anthropic is an AI safety company that builds and deploys large language models — the kind of AI that can read, write, reason, and write code. Its flagship product line is Claude, a family of AI assistants and APIs used by millions of people and integrated into tools like GitHub Copilot, Cursor, Excel, and PowerPoint. The company was founded with an explicit focus on making AI that is powerful and safe — and that combination has shaped nearly every major decision it has made.
Why it matters
Among the handful of companies defining what AI can do today, Anthropic stands out for two reasons. First, it consistently produces models at the very frontier of capability — its Claude Opus 4.8 topped major intelligence rankings, and its newer Claude Fable 5 and Mythos 5 set new records across software engineering, scientific reasoning, and cybersecurity. Second, it has been willing to pay a real price for its safety commitments: when the U.S. Department of War demanded Anthropic remove safeguards against fully autonomous weapons and mass domestic surveillance, the company publicly refused, got formally designated a national security supply-chain risk, and said it would fight the designation in court.
That's not a typical corporate move. It signals that Anthropic's safety commitments aren't just marketing — they have teeth, even when the pressure comes from the government.
How Claude works (the basics)
Claude is trained on large amounts of text and code, then refined to follow a set of principles — a process Anthropic calls Constitutional AI. The result is an assistant that can answer questions, write and debug software, analyze documents, and run multi-step tasks autonomously over long periods. Think of it like a very capable colleague who can work independently on a project for hours, checking in when it hits something it's unsure about.
Anthropic offers Claude in tiers: smaller, faster, cheaper models (Haiku) for everyday tasks; mid-range models (Sonnet) for more demanding work; and flagship models (Opus, and now Fable/Mythos) for the hardest problems. The most powerful tier — Mythos — is currently restricted to approved organizations because of its exceptional ability to find and exploit software vulnerabilities.
The safety-capability tension in practice
Anthropic's most powerful models have created genuine dilemmas. Claude Opus 4.6 found 22 real security vulnerabilities in Firefox's code in two weeks — 14 of them high-severity. Claude Mythos Preview autonomously discovered thousands of vulnerabilities in popular operating systems and browsers during testing. These capabilities are genuinely useful for defenders, but they're also dangerous in the wrong hands.
Anthropic's response has been to build tiered access: a public version (Fable 5) with safety classifiers that block or degrade responses on sensitive topics, and a restricted version (Mythos 5) available only to vetted organizations through a program called Project Glasswing. That program has expanded to 150 organizations across power, water, healthcare, and communications sectors, and has already found more than 10,000 high- or critical-severity security flaws in critical infrastructure code.
The company also disclosed that Claude Code was used in a real cyberattack: a Chinese state-sponsored group jailbroke it to autonomously attack roughly 30 organizations across tech, finance, and government in late 2025 — the first documented large-scale cyberattack run largely without human intervention. Anthropic detected and disrupted the campaign, then published the details to help the broader security community.
The government standoff
The most dramatic episode in Anthropic's recent history was its confrontation with the U.S. Department of War. The department wanted Claude available for "any lawful use," including fully autonomous weapons and mass domestic surveillance of Americans. Anthropic refused both exceptions, citing democratic values and the current reliability limits of AI systems.
The standoff escalated: the department formally designated Anthropic a supply-chain risk (a label previously applied only to foreign companies), and there were public threats of civil and criminal consequences. Anthropic committed to challenging the designation in court while continuing to support all other lawful national security uses at nominal cost during any transition.
Separately, in June 2026, the Commerce Department imposed export controls on Fable 5 and Mythos 5 after a jailbreak was demonstrated. Anthropic complied while publicly disputing the severity of the threat, noting that comparable models — including OpenAI's GPT-5.5 — could produce the same outputs. The controls were lifted within three weeks.
The business behind the mission
Anthropic has grown at a remarkable pace. It reported $14 billion in annualized revenue in early 2026, then $30 billion by April, then $47 billion by May — more than 10x annual growth for three consecutive years. Eight of the Fortune 10 are Claude customers. Claude Code alone generates over $2.5 billion in annualized revenue and accounts for an estimated 4% of all public GitHub commits worldwide.
To fuel that growth, Anthropic has locked in massive computing commitments: a 10-year, $100 billion-plus deal with Amazon for up to 5 gigawatts of compute; a multi-gigawatt TPU agreement with Google and Broadcom; a $30 billion Azure compute commitment with Microsoft; and access to SpaceX's Colossus data center with over 220,000 NVIDIA GPUs. It has also raised $30 billion (Series G) and $65 billion (Series H) in back-to-back funding rounds, reaching a $965 billion valuation.
High-profile talent has followed the money: AI pioneer Andrej Karpathy — co-founder of OpenAI and former head of Tesla Autopilot — joined Anthropic in mid-2026.
Where it's heading
Anthropic is expanding beyond text and code into scientific research: it announced Claude Science, an autonomous research agent for life sciences workflows, positioning it as the scientific equivalent of Claude Code. It also open-sourced the Model Context Protocol (MCP), a universal standard for connecting AI assistants to external tools and data sources, which has been adopted by companies like Block, Apollo, GitHub, and Slack.
The picture that emerges is a company trying to do something genuinely difficult: stay at the frontier of AI capability while maintaining limits that most of its competitors don't impose — and doing so at a scale that is rapidly approaching the size of major technology companies.
