Anthropic released Claude Opus 4.6 today, positioning the model for complex enterprise workflows in finance and security rather than general productivity gains.
The model's 1 million token context window (in beta) and 90.2% BigLaw Bench score represent the highest performance for Claude models on legal document analysis. In Rakuten's internal testing, Opus 4.6 autonomously closed 13 IT issues and assigned 12 others without human intervention. Security testing revealed 500+ unknown high-severity vulnerabilities across open-source libraries, including previously undetected flaws in GhostScript and OpenSC.
Anthropic added /effort controls (low to max) and adaptive thinking via API to manage agentic behavior. The model supports multi-document analysis across regulatory filings, market data, and company records. On code review benchmarks (MRCR v2 needle-in-haystack), Opus 4.6 scored 76% versus 18.5% for predecessors.
Pricing remains unchanged from Opus 4.5: $5 per million input tokens, $25 per million output tokens. Extended context (over 200K tokens) costs $10/$37.50 per million. The model ships through Anthropic's API, AWS Bedrock, and Google Vertex AI, with US-only inference options available. Anthropic says it now serves 300,000+ business users.
What's notable: Anthropic focused this release on specific enterprise use cases rather than benchmark chasing. The company added real-time blocking for malicious traffic patterns, signaling awareness that security research capabilities require safeguards.
The 128K maximum output tokens and enhanced agentic capabilities suggest Anthropic expects enterprises to use Opus 4.6 for entire workflow automation rather than single-query tasks. Whether that matches how organizations actually implement these models is the question that matters. Early adopters in legal, financial services, and security operations are the ones to watch.
Anthropic recommends migrating from earlier Claude 4 models (Sonnet 4.5, Opus 4.5) for intelligence gains, particularly on long-context tasks. The company's claim of "near-perfect technical scores" on internal benchmarks deserves scrutiny as independent testing expands.