Claude Opus 4.8

Overview. Claude Opus 4.8 is Anthropic's proprietary, closed-weight instruct model in the Claude Opus 4 family, accepting text and image input and producing text output over a 1M-token context window. It was Anthropic's flagship Opus model until Claude Opus 5 shipped in July 2026; Opus 4.8 remains active under Anthropic's 'legacy' status designation, now positioned as the established option behind the newer release. Developers reach for it when they need frontier-grade reasoning and agentic tool use rather than the newest model by default.

Key strengths. Opus 4.8 posts a GPQA-Diamond score of 0.92 and a tau2 (tool-agent-user) score of 0.94, indicating strong graduate-level reasoning and reliable structured tool-calling behaviour. Its Artificial Analysis intelligence index of 57.3 and coding index of 74.3 are the highest recorded across the Claude Opus 4 family, and Terminal-Bench v2.1 performance of 0.85 backs up its framing as a model built for complex agentic workflows. Long-context reasoning (0.73) supports its 1M-token window rather than treating it as a marketing number alone.

When to use it. Use Opus 4.8 for graduate-level analysis, multi-step agentic workflows, and tool-calling tasks where reasoning quality matters more than raw response speed. Its 1M-token context and long-context-reasoning score make it a fit for reviewing large documents or codebases in a single pass, and its opt-in adaptive thinking mode suits problems that benefit from a visible reasoning trace. It also accepts image input alongside text, so it can handle mixed-modality prompts within the same workflow.

Limitations. Opus 4.8 runs with a batch-oriented latency profile: time-to-first-token has ranged from roughly 7 to 38 seconds across recent measurements and throughput sits around 60 tokens per second, so it is not suited to low-latency, real-time interactions. Standard API output is capped at 128K tokens, and temperature, top_p, and top_k cannot be overridden at all — a constraint that holds regardless of thinking mode, with only opt-in adaptive thinking (not manual extended-thinking budgets) available on this model. Its training data cutoff is January 2026, and it has been superseded as Anthropic's flagship Opus model by Claude Opus 5.

Sovereignty note. Claude Opus 4.8 is built by Anthropic, a US-based creator, which means the model carries US CLOUD Act data-residency exposure rather than EU-Sovereign status. No findings indicate a training-on-input policy change or EU-resident endpoint option for this model.