On February 5, Anthropic released Opus 4.6, with the headline upgrade being Adaptive Thinking.
Before vs. Now
Previously, using extended thinking required manually setting a budget — developers had to decide whether the model should think for 10 seconds or 30 seconds. Now Opus 4.6 handles it on its own:
- Simple questions → instant response, no wasted compute
- Complex questions → automatically enters deep reasoning mode
- Compute resources allocated on demand
The underlying principle is that the model performs a difficulty assessment the moment it receives a prompt, then dynamically adjusts reasoning depth. Four manual levels are available (low/medium/high/max), with high as the default.
Context Window Maxed Out
- Context: 1 million tokens (beta)
- Maximum output: 128K tokens
- MRCR v2 retrieval accuracy: 93% at 256K context, still 76% at 1 million
For comparison, Sonnet 4.5's retrieval reliability under the same test is only one-quarter to one-ninth that of Opus 4.6. The gap is substantial.
Significant Agent Scenario Improvements
| Benchmark | Score |
|---|---|
| Terminal Bench | 59.8% → 65.4% |
| OSWorld | 66.3% → 72.7% |
Both scores surpass GPT-5.2 and Gemini 3 Pro.
Another standout feature is the Compaction API — server-side automatic compression of conversation history, theoretically supporting unlimited-length conversations. In agent scenarios that require dozens of interaction rounds to complete complex programming tasks, this capability directly determines task success rates.
Sources: The New Stack Opus 4.6 report, CocoLoop, Anthropic official documentation