Full deep dive: Claude Opus 4.6 Architecture Breakdown
Anthropic's flagship model, released February 2026. While the core architecture is proprietary, the published innovations — 1M token context window, adaptive thinking framework, and context compaction — reveal how Anthropic is building for long-running agentic workflows. Notable demo: 16 Opus 4.6 agents wrote a C compiler in Rust that compiles the Linux kernel.
What's Published
| Feature | Detail |
|---|---|
| Context window | 1M tokens (4x predecessor), 76% accuracy on MRCR v2 at 1M tokens |
| Output capacity | 128K tokens (doubled from 64K) |
| Adaptive thinking | Dynamic effort levels (low/medium/high/max) — model decides when to think |
| Context compaction | Auto-summarizes earlier conversation as context fills, combating "context rot" |
| Multimodal | Text, code, high-resolution images |
| Alignment | Constitutional AI (self-critique) |
Context Compaction
The most architecturally significant feature. As conversations approach the 1M limit, the API automatically summarizes earlier portions and replaces them with compressed state. This addresses "context rot" — the performance degradation that occurs as context windows fill — and is what makes the 1M window practically usable rather than just a marketing number.
Adaptive Thinking
Replaces static reasoning configurations with dynamic effort allocation. At the default (high) level, Claude almost always engages extended thinking. At lower levels, it may skip thinking entirely for simple tasks. This is the key to making a single model serve both quick Q&A and deep multi-step reasoning efficiently.
Benchmarks
| Benchmark | Score |
|---|---|
| Terminal-Bench 2.0 | 65.4% |
| OSWorld-Verified | 72.7% |
| ARC-AGI-2 | 68.8% |
| GPQA Diamond | 91.3% |
| Humanity's Last Exam (with tools) | 53.0% |
| METR 50%-time horizon | 14 hours 30 minutes (longest of any model) |
Why It's in Assess
Opus 4.6 is the model powering Claude Code and many of the agentic systems torn down elsewhere on this radar. The adaptive thinking and context compaction innovations are genuinely novel and worth understanding for anyone building long-running agent workflows. However, the proprietary architecture limits how much you can learn about how it achieves its results vs. open models like GLM-5 or GPT-OSS where you can inspect every layer. Assess the published patterns — especially context compaction and adaptive effort — as design principles for your own agent systems.
Key Characteristics
| Property | Value |
|---|---|
| Company | Anthropic |
| Model | Claude Opus 4.6 |
| Architecture | Transformer (proprietary enhancements) |
| Key innovations | 1M context, adaptive thinking, context compaction |
| Pricing | $5/$25 per M tokens |
| Available via | Anthropic API, AWS Bedrock, Google Vertex AI, Microsoft Copilot Studio |
| Released | February 5, 2026 |
| Sources | Anthropic Announcement, What's New, InfoQ |