Opus 4.6 is Anthropic’s strongest model for coding and long-running professional tasks. It is built for agents that operate across entire workflows rather than single prompts, making it especially effective for large codebases, complex refactors, and multi-step debugging that unfolds over time. The model shows deeper contextual understanding, stronger problem decomposition, and greater reliability on hard engineering tasks than prior generations.
Beyond coding, Opus 4.6 excels at sustained knowledge work. It produces near-production-ready documents, plans, and analyses in a single pass, and maintains coherence across very long outputs and extended sessions. This makes it a strong default for tasks that require persistence, judgment, and follow-through, such as technical design, migration planning, and end-to-end project execution.
Evaluations
33
across 20 benchmarks
Latency
1.3s
Context length
200k
tokens
Cost
$5 · $25
input · output per 1M tokens
Key takeaways
Claude Opus 4.6 is a Transformer-based model optimized for deep contextual understanding, problem decomposition, and reliability in demanding engineering and professional tasks. It is designed for agent-based workflows, large codebases, complex refactoring, and multi-step debugging.
This model excels at sustained knowledge work, producing near-production-ready documents and analyses. It maintains coherence over long outputs, making it ideal for technical design and end-to-end project execution.
The model achieved a perfect 100% accuracy on the 'AIME 2024 Math Problem-Solving Benchmark,' demonstrating strong abilities in advanced mathematical topics like algebra, geometry, number theory, and combinatorics. This highlights its robust mathematical reasoning and computational precision.
Limitations include struggles with precise quantitative tasks and mathematical errors (18.59% accuracy on 'Humanity's Last Exam'), misinterpretation of complex logic or subtle terminology, and inconsistent performance in specialized scientific domains.
Common failure modes in coding involve incorrect handling of edge cases, deviation from exact intended algorithms, and misinterpretation of specific problem constraints, as observed in tasks like `sort_third` and `is_nested`.
