Anthropic has quietly shipped one of the most developer-friendly features in frontier AI to date. Claude Opus 4.6 and Claude Sonnet 4.6 now support effort controls — a simple API parameter that lets builders dictate exactly how much cognitive horsepower the model applies to a given task.
What Are Effort Controls?
The new effort parameter replaces the older budget_tokens approach with four human-readable levels:
| Level | Use case |
|---|---|
low | Fast, cheap responses for simple lookups or routing |
medium | Balanced reasoning for standard tasks |
high | Default — deep chain-of-thought for complex problems |
max | Full extended thinking, no cap on reasoning tokens |
The model dynamically adjusts its internal chain-of-thought based on the selected level. Higher effort means more thorough reasoning but also higher latency and cost.
Why It’s a Big Deal
Previous extended thinking systems were all-or-nothing. You either paid for deep reasoning or you didn’t. Effort controls let developers match compute cost to task complexity — running low for thousands of cheap API calls while reserving max for the genuinely hard problems.
For agent frameworks and product builders, this is a meaningful efficiency lever.
What Else Came With Opus 4.6?
Beyond effort controls, the release ships:
- 1M-token context window (currently in beta)
- Improved coding and agentic capabilities
- Context compaction — automatically summarising long conversation histories to maintain performance without hitting limits
Anthropic says developers should experiment with effort levels for their specific domains, as the optimal setting varies widely by task type.
Source: anthropic.com, infoq.com