Extended Thinking was introduced in the Claude 3.7 era, allowing the model to generate a hidden chain of thought before answering. In 2026, Anthropic introduced extreme granularity for this feature.
xhigh Effort ParameterYou can now set the effort parameter to xhigh (Extra High) alongside the standard low, medium, and high.
{
"model": "claude-opus-4-7",
"thinking": {
"type": "enabled",
"effort": "xhigh"
},
"messages": [...]
}
xhighxhigh effort parameter allows Claude to consume up to 128,000 thinking tokens before generating an output. This can be extremely expensive. Always use budget caps in production.