Claude Opus 5

Overview. Claude Opus 5 is Anthropic's flagship closed-weight reasoning model, built for developers who need a large context window and adjustable reasoning depth in one endpoint. It handles a 1M-token context with up to 128K tokens of output, and accepts text, image, and file input. As a reasoning-tuned variant, it exposes a configurable effort ladder (low, medium, high, xhigh, max) so callers can trade latency for depth on a per-request basis.

Key strengths. Opus 5 sits in Stav's frontier quality tier, in line with the sibling Opus 4.x generation's positioning. Its complexity fit is reasoning: extended thinking is enabled by default and tunable through the effort ladder, and it accepts document (file) and image input. Domain signal estimates put code among its strongest areas, it reaches up to 94% on GPQA Diamond (graduate-level science reasoning) at the highest effort settings, and it shows solid general-purpose and writing performance.

When to use it. Reach for Opus 5 when a task benefits from deep, adjustable reasoning — multi-step problem solving, code generation, or research synthesis over long documents that need the full 1M-token context. Use lower effort settings for interactive, latency-sensitive chat, and dial up to high/xhigh/max when correctness matters more than response time, such as complex agentic tool-use chains or legal/technical document review. Its file and image input support also make it a fit for document-heavy or multimodal workflows.

Limitations. Anthropic has not published a training data cutoff or architecture details for Opus 5, so exact knowledge freshness and parameter scale can't be confirmed. Output is text-only — there is no native audio or image generation — and the model's true response latency varies significantly across the reasoning effort ladder, from interactive at low effort to considerably longer at max effort.

Sovereignty note. Claude Opus 5 is a closed-weight model from Anthropic, a US company subject to the CLOUD Act. It is available only as a hosted API, with no self-hosted or open-weight distribution. Where requests run, and whether inputs are retained or used for training, depends on the endpoint serving it; see the endpoint facts on this page.