Claude Opus 4.5

Claude Opus 4.5 is the flagship language model from Anthropic, released on November 24, 2025. The API model identifier is claude-opus-4-5-20251101. The model belongs to the Claude 4 family and served as Anthropic's most capable model until the release of Claude Opus 4.6.

Key features

Opus 4.5 introduces an effort parameter controlling reasoning intensity, extended thinking, advanced computer use, and a multi-tool calling mode. The context window is 200,000 tokens, with a maximum output of 64,000 tokens. Knowledge cutoff date: May 2025.

Benchmark results

Opus 4.5 is the first model to exceed the 80% threshold on SWE-bench Verified (80.9%), ahead of GPT-5.1 (76.3%) and Gemini 3 Pro (76.2%). It scores 59.3% on Terminal-Bench 2.0, 37.6% on ARC-AGI-2, 66.3% on OSWorld (a threefold improvement over Claude 3.5), and 87.0% on GPQA Diamond. On the multi-step tool use benchmark MCP Atlas, it achieves 62.3% — significantly above the second-place result (Claude Sonnet 4.5, 43.8%).

Safety

The model is deployed under the AI Safety Level 3 (ASL-3) standard. According to Anthropic, it is their best-aligned model at the time of release. It offers industry-leading resistance to prompt injection — 4.7% attack success rate (Gray Swan), compared to 12.5% for Gemini 3 Pro and 21.9% for GPT-5.1 (lower is better). Training incorporates RLHF, RLAIF (Constitutional AI), and an inoculation strategy against reward hacking.

Availability and pricing

The model uses closed weights and is available via the Claude API (Anthropic), Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing: $5 per million input tokens and $25 per million output tokens — a reduction of approximately 67% compared to Opus 4.1 ($15/$75).