Claude Sonnet 4.6: What's New for Architects

Claude Sonnet 4.6 brings several improvements that matter for architects building production systems. This article summarises the changes most relevant to agentic workflows, tool use, and the CCA-F exam domains.

Expanded Context Window

Sonnet 4.6 supports a 200,000-token context window (up from 100,000 in earlier Sonnet versions). For architects, this changes several Domain 5 decisions: tasks that previously required aggressive context management strategies can now be handled more simply with the larger window.

However, the exam's context management domain still tests the right patterns for when context does fill. Larger windows delay the problem; they don't eliminate the need to understand rolling summaries, anchoring, and RAG for truly long-running workflows.

Improved Tool Use Reliability

Sonnet 4.6 shows measurably better tool selection accuracy, especially in scenarios with many available tools. This is directly relevant to Domain 1: the practical threshold where tool selection reliability degrades has shifted slightly upward, but the architectural principle (keep tool counts low per agent; use specialisation) remains correct and is still tested.

Extended Thinking Improvements

Extended thinking — Claude's ability to reason step-by-step before producing a final answer — is more reliable in Sonnet 4.6 and better integrated with tool use. Agents can now run extended thinking between tool calls, not just before the first response.

For architects: this changes the cost model for complex agentic workflows. Extended thinking tokens are billed separately and can significantly increase cost per turn for complex reasoning tasks. Architects need to budget for thinking tokens explicitly.

Agent SDK Updates

The Claude Agent SDK received updates alongside Sonnet 4.6, primarily:

• Better error propagation from subagent failures to parent agents
• New lifecycle hooks for monitoring agent state across multi-turn sessions
• Improved handling of tool call retries with exponential backoff

What This Means for CCA-F Preparation

The CCA-F exam is based on architectural principles that apply across Claude model versions, not version-specific features. The core judgment traps (system prompt vs. hard gates, self-review, stop_reason vs. content analysis, etc.) remain stable. Model capability improvements don't change the correct architectural approach — they just change the baseline performance.