Anthropic has shipped Claude 4 Opus to general availability, bringing a 2-million-token context window, native multimodal reasoning, and an upgraded constitutional AI framework to its most powerful model.
The 2M context window — double that of Claude 3.5 Sonnet — allows Claude to process entire codebases, full-length books, or thousands of documents in a single conversation. In internal testing, the model maintains near-perfect recall (Needle In A Haystack score of 99.7%) even at the full 2M token length.
Native Multimodal Reasoning
Unlike previous versions that processed images and text separately, Claude 4 Opus reasons jointly across modalities. This means it can interpret charts, diagrams, and UI screenshots with the same fluency it reads text — a capability that proved critical in achieving top scores on visual reasoning benchmarks.
On MMMU (Massive Multi-discipline Multimodal Understanding), Claude 4 Opus scores 71.8%, surpassing GPT-5’s 69.2%. On document understanding tasks, it leads by an even wider margin.
Pricing
Claude 4 Opus costs $15/M input tokens and $75/M output tokens. The Claude 4 Sonnet model, released in beta, costs $3/$15 and handles most production workloads at 90% of Opus quality on standard tasks.
Anthropic says Claude 4 was trained using their new “iterative distillation” approach, which produces more predictable and interpretable outputs. The model also introduces a “thinking” mode that shows its reasoning chain before answering.