Anthropic released Claude 4 Sonnet and Claude 4 Opus in 2026, completing the generational upgrade it began with Claude 3. The two models serve meaningfully different use cases, and the distinction is sharper than in the Claude 3 generation.
Claude 4 Sonnet is the workhorse. It is the model Anthropic routes most API traffic through, and with good reason: benchmark performance on coding tasks (SWE-bench) sits at 72.7 percent for Sonnet, up from 49 percent for Claude 3.5 Sonnet. On MMLU, Sonnet 4 scores 90.2 percent. It handles 200,000-token contexts natively and can be extended to larger windows with document chunking. For the majority of production workloads — code completion, document analysis, customer support, structured data extraction — Sonnet 4 is the practical choice.
Claude 4 Opus is positioned for tasks where raw capability matters more than cost or speed. Its SWE-bench score is 72.5 percent — slightly below Sonnet in that specific benchmark, which surprised some observers — but it outperforms on multi-step reasoning tasks, complex instruction following, and sustained coherence across long documents. Anthropic says Opus is particularly strong on tasks that require judgment calls rather than pattern matching, such as legal document review or scientific literature synthesis. It is also slower and more expensive per token.
Anthropic's API pricing reflects the positioning. Sonnet 4 costs $3 per million input tokens and $15 per million output tokens. Opus 4 costs $15 per million input and $75 per million output — a 5x premium. That premium is substantial enough that most developers building at scale start with Sonnet and only route specific task categories to Opus when testing shows a meaningful quality difference.
The Anthropic Max subscription tier, which provides flat-rate access to both models for individual users, has become increasingly popular with developers who use Claude as a coding assistant during the day and want to experiment with long-context tasks without tracking per-token costs.