Claude 4
Claude 4
Claude 4 is Anthropic's frontier model family, released in stages throughout 2025-2026. It includes Claude 4 Opus (flagship), Claude 4 Sonnet (balanced), and Claude 4 Haiku (fast/light).
Key Specifications
- Context window: 200K tokens (standard), 500K (Opus extended)
- Max output tokens: 8,192
- Architecture: Transformer (proprietary)
- Modalities: Text, image input; text output
- Training: Extensive RLHF with constitutional AI principles
Model Variants
| Variant | Description | Pricing (Input/Output per 1M) |
|---|---|---|
| Opus | Flagship, best reasoning | $15.00 / $75.00 |
| Sonnet | Balanced quality/speed | $3.00 / $15.00 |
| Haiku | Fastest, most affordable | $0.25 / $1.25 |
Capabilities
- Extended reasoning — Multi-step chain-of-thought reasoning
- Code generation — Strong at complex coding tasks
- Vision — Document analysis, image understanding
- Tool use — Native function calling
- Long context — Handles book-length documents
- Agentic — Can autonomously execute multi-step tasks
Safety Approach
Claude 4 uses constitutional-ai with a "harmlessness" training objective alongside helpfulness. Anthropic publishes detailed model cards and safety evaluations. The model includes a "thinking" mode for transparent reasoning.
Comparison
- GPT-4o — Primary closed-source competitor
- Gemini 2 Pro — Google's competing model
- See reasoning-models-comparison for reasoning benchmarks