What Claude is
Claude is Anthropic's family of large language models — the commercial and research expression of the company's bet that frontier AI capability and rigorous safety practice can be built together. The lineup is tiered: the flagship Opus 4.8 handles the most demanding reasoning, coding, and long-horizon agentic tasks; Sonnet 4.6 targets balanced speed and quality for enterprise workflows; Haiku 4.5 serves high-volume, low-latency applications; and Fable 5 is a dedicated creative-tier model. All tiers share the same Constitutional AI (CAI) training foundation and are governed by Anthropic's Responsible Scaling Policy (RSP).
Architecture and training approach
Anthropic does not publicly disclose Claude's internal architecture. What is documented is the training methodology: Constitutional AI uses a written "constitution" — drawing from sources including the UN Declaration of Human Rights, DeepMind's Sparrow Principles, and Anthropic's own safety research — to generate AI-guided self-critique and reinforcement learning signals, rather than relying exclusively on large-scale human feedback. The constitution establishes a priority hierarchy: broadly safe → broadly ethical → compliant with Anthropic guidelines → genuinely helpful. The January 2026 revision of Claude's Constitution was released under CC0, allowing unrestricted reuse.
The RSP formalizes risk management through AI Safety Levels (ASL-1 through ASL-5+), modeled on U.S. biosafety standards. Current Claude models are classified ASL-2; ASL-3 triggers mandatory adversarial red-teaming and stricter deployment constraints. Anthropic's alignment research, published in 2026, showed that training Claude on ethical reasoning (rather than just aligned actions) reduced agentic misalignment rates from 22% to 3%, with every model from Haiku 4.5 onward scoring perfectly on misalignment evaluations.
Capability trajectory
Claude's public history traces a rapid capability arc. The March 2023 public launch introduced two tiers (Claude and Claude Instant) with early partners reporting fewer harmful outputs and better steerability than competing models. A context window expansion from 9K to 100K tokens followed in May 2023, enabling full-codebase and multi-document processing. Claude 2 (mid-2024) pushed coding performance to 71.2% on Codex HumanEval and 88.0% on GSM8k math. Claude 2.1 introduced the first 200K-token context window and a tool-use beta. The Claude 4.x family extended these gains into agentic workflows, with Claude Code becoming a significant standalone product — generating over $2.5B in annualized run-rate revenue and accounting for an estimated 4% of all GitHub public commits worldwide by early 2026.
Deployment footprint
Claude is available across three major cloud platforms: Amazon Bedrock (primary training and inference partner, backed by a 10-year, $100B+ AWS commitment securing up to 5GW of Trainium2–4 compute), Google Cloud Vertex AI (up to one million TPUs in a tens-of-billions deal, with multi-gigawatt next-generation TPU capacity coming online from 2027), and Microsoft Azure Foundry (backed by a $30B Azure compute commitment, with Claude models available across the Copilot product family).
Enterprise reach is substantial: eight of the Fortune 10 are Claude customers, over 1,000 businesses spend more than $1M annually, and major professional services firms have embedded Claude at scale — Deloitte (470,000 professionals), KPMG (276,000), PwC (hundreds of thousands), and Accenture (30,000 trained professionals in a dedicated Anthropic Business Group). Vertical deployments span financial services (Bridgewater, RBC Wealth Management), healthcare and life sciences (ServiceNow's claims authorization workflows, HHMI's Janelia Research Campus), government (GOV.UK job-seeker assistant, U.S. Department of Energy Genesis Mission across 17 national laboratories), and scientific research (Allen Institute multi-agent biological data analysis).
Claude also powers Amazon's Alexa+, extending its reach to Amazon's consumer device footprint.
The governance flashpoint
Claude's safety posture has produced the most consequential regulatory confrontation in U.S. AI history to date. Anthropic maintains two hard usage exceptions: fully autonomous weapons and mass domestic surveillance. When the U.S. Department of War demanded Anthropic remove these restrictions and accept "any lawful use," Anthropic refused — publicly, in a statement from CEO Dario Amodei. The DoW responded by designating Anthropic a supply-chain risk under 10 USC 3252, a designation previously applied only to foreign companies. Anthropic committed to challenging the designation in court while continuing to provide models to the national security community at nominal cost during any transition. OpenAI signed a contract allowing use of its models "for all lawful purposes," though that agreement was later described as rushed and renegotiated.
The episode was complicated by a concurrent revelation: Claude, integrated with Palantir's Maven Smart System, was used to accelerate U.S. military targeting in Iran — reportedly compressing a 12-hour targeting process to under one minute and helping select over 1,000 targets in the first 24 hours of operations. A subsequent investigation found U.S. forces likely struck a school killing 170+ people, with stale target data cited as a potential contributing factor. Iranian drone strikes on AWS data centers in Bahrain and the UAE followed. The episode marks the first known large-scale operational use of a commercial LLM in active kinetic conflict.
Adversarial threats and misuse
Claude's differentiated capabilities have made it a target. Anthropic identified three Chinese AI laboratories — DeepSeek, Moonshot AI, and MiniMax — as conducting coordinated industrial-scale distillation attacks: over 16 million exchanges through approximately 24,000 fraudulent accounts, targeting agentic reasoning, tool use, coding, and chain-of-thought generation. MiniMax alone was responsible for over 13 million exchanges. Anthropic frames these as a national security concern, arguing distilled models strip out safety safeguards and undermine U.S. export controls.
On the offensive use side, Anthropic detected and disrupted a Chinese state-sponsored cyberattack in September 2025 that used Claude Code as an autonomous agent against roughly thirty global targets — the first documented large-scale cyberattack executed without substantial human intervention. Earlier misuse reports documented influence operations (100+ social media bots with Claude as agentic orchestrator), data extortion campaigns, ransomware development, and North Korean fraudulent employment schemes. Anthropic co-developed a nuclear safeguards classifier with the NNSA achieving 96% accuracy on live Claude traffic, shared with the Frontier Model Forum as a replicable blueprint.
Commercial scale and infrastructure
Anthropic's revenue trajectory has been steep: from approximately $1B annualized at the start of 2025, to $9B by end of 2025, to $30B+ by April 2026, to $47B by May 2026. The company closed a $65B Series H at a $965B post-money valuation in May 2026, following a $30B Series G at $380B in February 2026 and a $13B Series F at $183B in November 2025. Compute commitments span Amazon (up to 5GW Trainium), Google/Broadcom (5GW TPU), Microsoft/NVIDIA (up to 1GW Grace Blackwell/Vera Rubin), and a $50B domestic infrastructure commitment with Fluidstack building custom data centers in Texas and New York.
Where it's heading
The events in this bundle point toward three simultaneous expansions: deeper agentic capability (multi-agent harnesses, Claude Code as autonomous software engineer, MCP-connected tool ecosystems), broader enterprise infrastructure (a new AI services company co-founded with Blackstone, Hellman & Friedman, and Goldman Sachs targeting mid-market enterprises), and continued navigation of the hardest governance questions in the field — where the line sits between lawful national security use and uses Claude's constitution prohibits, and how to defend differentiated capabilities against adversarial extraction at industrial scale.
