Anthropic releases Claude with extended reasoning + agent SDK improvements

The release adds finer-grained control over chain-of-thought reasoning visibility, expands tool-use guarantees for parallel calls, and ships a refreshed Agent SDK for production agent deployments. The headline improvement is the extended thinking mode — Claude can now reason for up to 200,000 tokens before responding, with explicit budget controls so developers can cap thinking time for latency-sensitive applications.

Three changes matter practically for enterprise teams. First, **parallel tool calls**: Claude can now invoke multiple tools simultaneously in a single turn, reducing round-trips in agentic workflows by 40-60% on multi-step tasks. This directly addresses the latency bottleneck that made Claude less competitive than GPT-4o in production agent systems. Second, **reasoning visibility controls**: enterprises can choose to expose the chain-of-thought to end users (for compliance and explainability) or suppress it (for cost control, since thinking tokens are charged at output rates). Third, **SDK stability guarantees**: Anthropic has committed to maintaining agent SDK interfaces across minor version bumps — an important signal for teams building production systems that cannot absorb constant migration overhead.

For enterprise teams already on Claude, the migration path is straightforward — most existing prompt structures work without modification. The cost calculus shifts: extended thinking mode costs more per token but may reduce total cost if it eliminates agentic retry loops caused by reasoning failures. AIMenta recommends running a cost-per-successful-task benchmark across your top five workflows before deciding whether to enable extended thinking by default.

The Agent SDK improvements are the most strategically significant change for APAC clients building multi-step document processing, data extraction, or customer service orchestration. These patterns, previously requiring careful scaffolding to avoid tool-call failures, are now more reliable out of the box. Teams that deprioritised Claude for agent work due to reliability concerns should re-evaluate.