Claude (Language Model)

Claude is a family of large language models developed by Anthropic, an American AI safety company founded in 2021 by former OpenAI researchers including Dario and Daniela Amodei. The name is widely understood to honour Claude Shannon, the founder of information theory. Claude models are designed with a strong emphasis on safety, interpretability, and adherence to a written set of guiding principles using a training method Anthropic calls Constitutional AI.

History and generations

Anthropic released the first Claude model in March 2023, followed by Claude 2 in July 2023, which introduced a 100,000-token context window — a significant leap over contemporary models. The Claude 3 family, released in March 2024, established a three-tier naming convention that has continued through later generations: Haiku (smallest and fastest), Sonnet (balanced), and Opus (most capable).

Subsequent generations refined the family. Claude 3.5 Sonnet and Claude 3.7 Sonnet introduced extended thinking, native tool use, and improved coding ability. The Claude 4 generation, released across 2025 and 2026, brought multi-hour autonomous task execution, sharper agentic behaviour, and stronger software engineering performance. Claude Sonnet 4.5 achieved 77.2% on the SWE-bench Verified coding benchmark and demonstrated the ability to maintain coherent work on engineering tasks across more than thirty hours.

Architecture and training

Claude models are transformer-based decoder-only language models trained on a mixture of publicly available text, licensed data, and human feedback data. Anthropic has not published full parameter counts for production Claude models, but they are widely understood to range from billions to hundreds of billions of parameters across the family tiers.

Claude is trained using a combination of pre-training on large text corpora, supervised fine-tuning, reinforcement learning from human feedback (RLHF), and Anthropic's proprietary Constitutional AI procedure. Constitutional AI uses a written list of principles — drawn from sources such as the United Nations Universal Declaration of Human Rights and Apple's Terms of Service — to guide an AI feedback model that critiques and revises the model's outputs during training. This reduces the volume of human labelling required to align the model with desired behaviours.

Capabilities

All current Claude models accept text and image inputs and produce text outputs, with native support for tool use, JSON-mode structured outputs, and long-context retrieval. The family supports context windows ranging from 200,000 tokens in the standard tier up to 1 million tokens in extended-context offerings, allowing entire codebases or document collections to be processed in a single request.

Claude is widely used for software engineering, legal and financial analysis, scientific research summarisation, customer support automation, and the construction of agentic workflows where the model plans and executes multi-step tasks with tool use. Anthropic also distributes Claude through the Claude.ai web product, the Claude Code command-line tool, and through cloud partners including Amazon Bedrock and Google Vertex AI.

Safety research

Anthropic positions safety research as central to its product strategy. The company publishes work on mechanistic interpretability, red-teaming, jailbreak resistance, and model evaluations under its Responsible Scaling Policy. Claude models are subject to a tiered AI Safety Level (ASL) framework that mandates additional safeguards as capabilities grow.

See Also

References

References

1. Anthropic. (2024). Introducing the Claude 3 Family. Anthropic Research.
2. Bai, Y., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.
3. Anthropic. (2025). Responsible Scaling Policy v2. Anthropic Policy Documents.
4. Bank Negara Malaysia. (2023). Risk Management in Technology (RMiT) Policy Document.
5. Malaysia Digital Economy Corporation. (2024). National AI Roadmap and Industry Engagement Reports. MDEC.