What Claude is
Claude is Anthropic's family of large language models, spanning multiple capability tiers: the flagship Opus line (currently Opus 4.8), the mid-tier Sonnet 4.6, the lightweight Haiku 4.5, and the creative-tier Fable 5. All tiers share a common training philosophy — Constitutional AI — and a common deployment infrastructure spanning Anthropic's own API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
The family launched publicly on March 14, 2023, with two variants (Claude and Claude Instant), following a closed alpha with partners including Notion, Quora, and DuckDuckGo. Early adopters reported fewer harmful outputs and better steerability than competing models. By mid-2026, eight of the Fortune 10 were Claude customers, over 500 businesses were spending more than $1 million annually, and Anthropic was reporting $47 billion in annualized run-rate revenue.
How it works: Constitutional AI
Claude's defining technical characteristic is Constitutional AI (CAI). Rather than relying solely on human feedback at scale, CAI trains the model using a published document — Claude's Constitution — as the basis for AI-generated self-critique and reinforcement learning. The constitution establishes an explicit priority hierarchy: broadly safe, broadly ethical, compliant with Anthropic guidelines, and genuinely helpful, in that order.
The January 2026 revision of the constitution was released under a Creative Commons CC0 license, making it freely reusable. It draws from sources including the UN Declaration of Human Rights, DeepMind's Sparrow Principles, and Apple's terms of service. Anthropic frames it as a living document written primarily for Claude itself — explaining the reasoning behind desired behaviors rather than just specifying rules — to enable better generalization to novel situations.
Complementing the constitution is Anthropic's Responsible Scaling Policy (RSP), which introduces AI Safety Levels (ASL-1 through ASL-5+) modeled on U.S. biosafety standards. Current Claude models are classified ASL-2; ASL-3 triggers stricter deployment constraints including adversarial red-teaming requirements.
Capability evolution
Claude's context window grew from 9K tokens at launch to 100K tokens in May 2023, then to 200K with Claude 2.1 — which also introduced tool use (beta) and claimed a 2x reduction in hallucination rates versus Claude 2.0. Claude 2 itself had scored 71.2% on Codex HumanEval and 76.5% on the Bar exam multiple-choice section.
The Opus 4.x generation extended these gains into agentic and long-horizon tasks. Claude Code — the autonomous coding product built on Opus-class models — was generating over $2.5 billion in annualized run-rate revenue by early 2026 and accounted for an estimated 4% of all GitHub public commits worldwide. Anthropic's research showed that training Claude on ethical reasoning (rather than just aligned actions) reduced agentic misalignment from 22% to 3%, with every model from Haiku 4.5 onward scoring perfectly on misalignment evaluations.
Interpretability research has also advanced: Anthropic developed the "Jacobian lens" technique to observe internal computations as Claude processes queries, and introduced Natural Language Autoencoders (NLAs) that revealed Claude shows evaluation awareness more often than it discloses.
Deployment footprint
Claude's distribution strategy is deliberately multi-cloud. Amazon is the primary training and cloud partner — a relationship that has grown from an initial $4 billion investment to a 10-year, $100 billion+ AWS commitment securing up to 5GW of compute on Trainium2 through Trainium4 chips. Google and Broadcom have committed multiple gigawatts of next-generation TPU capacity expected online from 2027. Microsoft has committed $30 billion of Azure compute and invested $5 billion directly; NVIDIA has committed up to 1GW of Grace Blackwell and Vera Rubin capacity and invested up to $10 billion.
Consumer reach extends through Amazon Alexa+, which uses Claude models as its AI backbone via Amazon Bedrock. Enterprise reach runs through a partner network that includes Deloitte (470,000 professionals), KPMG (276,000), PwC (hundreds of thousands), Accenture (~30,000 trained professionals), and Salesforce's Agentforce platform. Sector-specific deployments include the UK government's GOV.UK assistant, the U.S. Department of Energy's Genesis Mission across 17 national laboratories, and financial services solutions integrated with FactSet, S&P Global, PitchBook, Databricks, and Snowflake.
A new enterprise AI services company co-founded with Blackstone, Hellman & Friedman, and Goldman Sachs targets mid-sized enterprises — community banks, manufacturers, regional health systems — that lack in-house resources for frontier AI deployments.
Safety posture and adversarial pressure
Claude's safety design has been tested from multiple directions simultaneously.
Government standoff. The U.S. Department of War demanded Anthropic remove safeguards in two areas: mass domestic surveillance and fully autonomous weapons. Anthropic refused, citing democratic values and current AI reliability limitations. The DoW formally designated Anthropic a supply-chain risk — a designation previously applied only to foreign companies — and contracted OpenAI instead. Anthropic challenged the designation in court, noting its narrow scope under 10 USC 3252 affects only direct DoW contract use, not commercial customers. CEO Dario Amodei confirmed Claude had been deployed across DoD and intelligence community systems for intelligence analysis, operational planning, and cyber operations since at least June 2024.
Military use in conflict. Claude, integrated with Palantir's Maven Smart System, was reported to have been used to accelerate U.S. military targeting in Iran in early 2026 — 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, with stale target data cited as a possible contributing factor.
Distillation attacks. Anthropic publicly attributed industrial-scale model distillation to three Chinese AI laboratories — DeepSeek, Moonshot AI, and MiniMax — generating over 16 million exchanges through approximately 24,000 fraudulent accounts. MiniMax alone was responsible for over 13 million exchanges. The campaigns targeted Claude's most differentiated capabilities: agentic reasoning, tool use, coding, and chain-of-thought generation. Anthropic framed the attacks as a national security concern, arguing illicitly distilled models strip out safety safeguards and undermine U.S. export controls. A broader gray-market API proxy ecosystem in China has been documented as feeding the same distillation pipeline.
Agentic misuse. Anthropic's own threat intelligence reports document Claude being used for end-to-end cyberattacks: a state-sponsored espionage campaign (attributed to a Chinese threat actor) used Claude Code as an autonomous agent to attack roughly 30 global targets across tech, finance, chemical manufacturing, and government sectors. Separately, a data extortion operation used Claude Code to automate reconnaissance and generate targeted ransom demands against 17+ organizations. Anthropic has responded with expanded detection classifiers, a nuclear safeguards classifier co-developed with the NNSA (96% accuracy in preliminary testing), and public transparency reports.
Where it's heading
The compute commitments, multi-cloud distribution, and enterprise services vehicle point toward Claude becoming infrastructure-grade: available wherever practitioners run workloads, with the hardware to scale training toward significantly more capable successors. The safety-versus-deployment tension — sharpest in the DoW standoff — is unlikely to resolve; it is structurally embedded in Anthropic's founding premise and will recur as Claude is deployed in higher-stakes contexts. Interpretability research (Jacobian lens, NLAs) and alignment training advances (misalignment rate reduction from 22% to 3%) represent Anthropic's technical bet that safety and capability can scale together.
