A reasoning mechanism introduced in <a href="/en/glossary/claude-models/claude-opus/">Claude Opus</a> 4.8 and Sonnet 4.6 — the model automatically determines how much reasoning resource to invest based on question difficulty, requiring no user parameter settings. Simple questions get quick answers; complex questions automatically receive deeper reasoning. Unlike <a href="/en/glossary/core-concepts/extended-thinking/">Extended Thinking</a> (which requires explicit activation), Adaptive Thinking is an always-on background capability, completely transparent to users.

An AI system capable of autonomously executing multi-step tasks — not just answering questions, but actively planning, calling tools, observing results, and adjusting actions until a task is complete. The key difference from ordinary question-answering AI: it can decide its own next steps without human instruction at each stage.

The company behind Claude, founded in 2021 by former OpenAI researchers. Its core mission is developing safe and beneficial AI. Anthropic's safety research directly shapes Claude's design philosophy and behavioral characteristics.

An AI training method proposed by <a href="/en/glossary/core-concepts/anthropic/">Anthropic</a> that trains the model to self-evaluate and improve according to an explicit set of behavioral principles (the "constitution"), rather than relying entirely on human subjective scoring of individual responses. One of the core training mechanisms behind Claude's honesty and safety.

The engineering practice of systematically managing input <a href="/en/glossary/core-concepts/token/">Token</a> counts in API calls, with the goal of reducing costs and latency while maintaining output quality. Combines conversation history truncation, summary compression, <a href="/en/glossary/prompt-techniques/system-prompt/">System Prompt</a> trimming, and related techniques. An essential skill for production <a href="/en/glossary/claude-tools/claude-api/">Claude API</a> deployments.

The maximum amount of text an AI model can "see" and process in a single conversation. When content exceeds this limit, the earliest content gets pushed out of the window and the model "forgets" that part of the conversation. Claude's Context Window is 200,000 tokens — roughly 150,000 English words or 100,000 Chinese characters, about the length of a thick novel.

<a href="/en/glossary/workspace-basics/extended-thinking/" target="_blank">Extended Thinking</a> is Claude's ability to work through reasoning in a hidden scratchpad before delivering its final answer. You set a thinking budget (a <a href="/en/glossary/core-concepts/token/">Token</a> ceiling), and Claude reasons freely within it — exploring, backtracking — before producing the output you see. The thinking itself is usually not visible, but it lets Claude handle complex problems with deeper, more structured reasoning, like working out scratch paper before speaking.

Further training a pre-trained language model on domain-specific data to make its behavior better suited to specific needs. The core technique for turning a generalist AI into a domain specialist.

The phenomenon where AI models generate information that seems plausible but is actually incorrect or fabricated. The name comes from human psychology: just as people "see" things that don't exist in dreams, AI "says" things it has no factual basis for — names, dates, citations, statistics — all potentially "sounds correct but doesn't actually exist" content.

A collection of techniques for reducing computational cost and latency during the AI model's "generating a response" (inference) stage. Unlike training optimization, inference optimization focuses on how an already-trained model can output results faster and more cheaply. Includes techniques such as quantization, batching, and speculative decoding.

The last date at which an AI model's training data was updated. Events after this date are completely unknown to the model. Claude's knowledge cutoff is end of August 2025 — asking about post-cutoff events may yield outdated or incorrect information.

A data structure using 'nodes (entities) + edges (relationships)' graph structure to represent and store knowledge. Unlike vector databases (similarity search), knowledge graphs preserve explicit relationships and logical structure between information. In AI applications, knowledge graphs typically pair with RAG — after vector search finds relevant content, the knowledge graph provides structured relationships between entities, enabling Claude to perform more precise multi-hop reasoning.

An AI model trained on massive text data with the ability to understand and generate human language. Claude, GPT-4, and Gemini are all LLMs. "Large" refers to the enormous number of model parameters, not the length of text output.

Using the outputs of a large "teacher" model to train a smaller "student" model, allowing the small model to retain the large model's core capabilities while dramatically reducing computational requirements. Like having a senior expert intensively mentor a junior colleague, compressing years of tacit knowledge — the junior reaches 80% of the expert's capability at just 10% of the "size."

An AI's ability to simultaneously process and understand multiple input types, including text, images, and documents. Claude's multimodal capabilities let you upload screenshots, photos, and PDFs to ask questions directly — no need to convert everything to text first.

An <a href="/en/glossary/core-concepts/anthropic/">Anthropic</a> API feature letting frequently reused prompt prefixes (like fixed System Prompts, long background documents) be computed only once; subsequent calls read the cached version at 90% lower cost and reduced latency. For applications with long fixed System Prompts, enabling Prompt Caching typically immediately saves 20-40% of total API costs.

An architecture that lets AI retrieve relevant information from an external knowledge base before generating a response — solving <a href="/en/glossary/core-concepts/knowledge-cutoff/">Knowledge Cutoff</a> and <a href="/en/glossary/core-concepts/hallucination/">Hallucination</a> problems.

A parameter controlling the "randomness" or "creativity level" of AI output, typically between 0 and 1. 0 makes the model give nearly identical, most certain answers every time (good for code generation, fact queries); 1 makes output more varied and creative but less predictable (good for creative writing). Intuitive analogy: low <a href="/en/glossary/prompt-techniques/temperature/" target="_blank">Temperature</a> is a careful analyst who only says what they're most confident about; high temperature is an inspired creator who voices more unusual ideas.

The smallest unit AI language models use to process text. Tokens don't equal words — English averages roughly 0.75 words per token; Chinese roughly 1-2 characters per token. API costs are billed per token: how many you input plus how many Claude outputs equals the cost of this conversation. Understanding tokens helps you more effectively control AI costs and Context Window usage efficiency.