Claude Opus 4.6

Overview. Claude Opus 4.6 is a proprietary, closed-weight model in Anthropic's Claude Opus 4 family, released in February 2026 — Anthropic has since released the higher-scoring Claude Opus 4.7 and Opus 4.8 within the same family. It pairs a one-million-token context window with a 128,000-token maximum output, and accepts text, images, and native PDF/document input alongside text output. Developers reach for it when a task calls for Anthropic's high-reasoning Opus tier, rather than the fastest or most economical option.

Key strengths. Opus 4.6 posted an Elo of 1497 on the LMSys Chatbot Arena leaderboard (ranked fourth as of Stav's last verified check), placing it in Stav's frontier quality tier. On Artificial Analysis's GPQA Diamond benchmark it scores 0.84, and on Humanity's Last Exam it scores 0.19 — strong absolute results for a general-purpose model. It also supports optional adaptive thinking with selectable effort levels (max, high, medium, low), letting a caller trade latency for deeper reasoning on demand.

When to use it. Use Opus 4.6 for reasoning-heavy and agentic workflows where output quality matters more than raw speed — complex multi-step tasks, long-document analysis, and research-style question answering all benefit from its full context window and adaptive thinking. Its native document and image understanding also make it a good fit for prompts that mix PDFs, screenshots, and text. With a median time-to-first-token around 2.4 seconds, it suits interactive use comfortably, though it isn't tuned for sub-second, high-throughput serving.

Limitations. Opus 4.6 is closed-weight and served only through Anthropic's own API and partner cloud platforms, so self-hosting or fine-tuning the base weights isn't possible. Its training data reaches through roughly August 2025, and it does not accept or produce audio.

Sovereignty note. Opus 4.6 is a proprietary, closed-weight model governed by Anthropic's Acceptable Use Policy and is available only via Anthropic's own hosted infrastructure or partner clouds, so data handling and residency depend on Anthropic's terms rather than a self-hosted deployment.