Textbook of AI

Claude 4 is Anthropic's late-2025 model family, succeeding Claude 3 (March 2024) and the Claude 3.5 / 3.7 interim releases. It comprises three publicly named tiers: Opus 4 (largest, highest-capability), Sonnet 4 (balanced workhorse) and Haiku 4 (small, fast, cheap). A 1M-token context variant of Opus 4 was released to address long-context engineering and research workloads.

Capability profile. Claude 4 advanced sharply on coding, agentic tool use and long-horizon reasoning. On SWE-Bench Verified Sonnet 4 and Opus 4 set state-of-the-art scores in their respective tiers, and the family became the default backbone of Claude Code, Anthropic's terminal-based software-engineering agent. Long-context retrieval at multi-hundred-thousand-token depth improved markedly compared to Claude 3.5, narrowing the gap with Gemini's million-token native window.

Reasoning. Claude 4 introduced extended thinking, a user-visible reasoning mode in which the model produces a structured internal scratchpad before answering. Unlike OpenAI's hidden o-series chains, Anthropic exposes summary thinking traces by default, citing interpretability and trust as design goals. Extended thinking is selectable per request and trades latency for accuracy on hard tasks.

Constitutional AI maturation. The training pipeline integrated several years of constitutional AI research: a written constitution of behavioural principles, AI feedback loops to flag and revise unsafe completions, character training to give the model a consistent persona, and classifier-based runtime safety filters. Anthropic published a "Claude's character" document and a system prompt specification, making the deployed behaviour more legible.

Multimodality. All three tiers accept images and PDFs natively. Audio input and richer document grounding shipped through 2025. Claude 4 does not generate images.

Agentic features. Claude 4 ships with mature support for the Model Context Protocol, parallel tool calls, and long-running sub-agent orchestration. The model's tool-use reliability is a major selling point: it issues fewer malformed calls and recovers more cleanly from tool errors than its predecessors, which is essential for multi-hour autonomous coding sessions.

Distribution. Available through claude.ai, the Anthropic API, AWS Bedrock, Google Vertex AI, and Microsoft Foundry. Anthropic's own Claude Code and Claude Desktop are the canonical first-party clients.

Position in the field. As of early 2026 Claude 4 Opus is regarded as one of the three frontier model families alongside GPT-5 and Gemini 2, with particular strength in coding, careful reasoning, and structured agent workflows. The release also marked Anthropic's clearest product-market fit: developer-tool usage on Claude Code drove a substantial share of API revenue.