Did Sonnet Just Beat Opus? — Claude Sonnet 4.6 Benchmark Deep Dive

Anthropic released Claude Sonnet 4.6 on February 17, and it outperforms the flagship Opus 4.6 on several key benchmarks. At roughly 40% less cost. The secret isn't a "cheaper knock-off" — it's architectural-level structural changes.

Opus vs Sonnet: What Changed?

The old Opus-Sonnet dynamic was straightforward. Opus was the full-spec brain; Sonnet was the compressed version. Same architecture, smaller size, naturally lower performance.

In the 4.6 generation, that formula breaks.

Where Sonnet Wins or Ties

In coding and agentic tasks, Sonnet matches or beats Opus. At $3/$15 per M tokens.

Where Opus Clearly Wins

See the pattern? Opus decisively wins on reasoning depth and ultra-long context accuracy.

Here's an analogy: Sonnet 4.6 is a student who scores perfect on the SAT. Within defined boundaries, it executes near-flawlessly. Opus 4.6 is an International Math Olympiad gold medalist. The gap emerges on novel problems that require chaining multiple concepts.

Most real-world coding, document work, and agentic tasks fall within "SAT territory." Opus is only necessary for research-grade tasks requiring novel reasoning.

The Evolution of Computer Use

The individual benchmark numbers aren't the real story.

Look at the OSWorld benchmark trajectory:

5x improvement in 16 months. Roughly 10-15 percentage points every three months.

If this curve holds, it crosses 90% within the year. "AI operating a computer like a human" becomes reality in 2026. Mouse clicks, drag-and-drop, form filling, file management — all performed directly by AI.

Adaptive Thinking: Automatic Reasoning Depth Control

Previous Extended Thinking always went deep. Even simple questions consumed tokens through excessive reasoning.

Adaptive Thinking auto-adjusts across 4 levels: low, medium, high, and max.

It's like how humans adjust thinking time based on problem difficulty. Same quality output, fewer tokens burned.

For developers, the key insight: the budget_tokens parameter lets you cap reasoning per request — "think this much and no more." Cost predictability is now possible.

Context Compaction: The Most Underrated Feature

The 1M token context window existed before, but the problem was "Lost-in-the-middle." Feeding a million tokens is pointless if the model forgets what's in the middle.

Context Compaction automatically summarizes older context server-side. It preserves key information while saving tokens.

Why does this matter? It could fundamentally change RAG pipeline design.

• Before: Document -> Chunking -> Embedding -> Vector DB -> Reranking -> LLM (5-stage pipeline)
• With 4.6: Document -> LLM (1 stage, Compaction handles the rest)

Of course, 1M context is still in beta and only accessible at Usage Tier 4+. Opus scored 76% on MRCR v2, but Sonnet 4.6's score hasn't been published yet. This part needs verification.

When Opus, When Sonnet?

Synthesizing these benchmarks, the conclusion is clear.

Default to Sonnet 4.6

• Coding, debugging, code review (SWE-bench 79.6%)
• Data analysis, documentation, knowledge work (GDPval-AA 1633 Elo)
• Agent workflows, tool use (Finance Agent 63.3%)
• Direct computer operation (OSWorld 72.5%)
• Price: $3 input / $15 output per M tokens

Use Opus 4.6 Only When

• Solving truly novel problems (ARC-AGI-2 68.8%)
• Olympiad-grade reasoning required (HLE 53.0%)
• Finding needles in million-token haystacks (MRCR v2 76%)
• Price: $5 input / $25 output per M tokens

If you're a developer, switch your default stack to Sonnet 4.6 and reserve Opus for reasoning-heavy nodes only. Same budget, ~1.67x more throughput.

Key Takeaways

Sonnet 4.6 isn't a "budget Opus." It's an optimized-for-production model. It matches or exceeds Opus on coding, agentic tasks, and knowledge work — at ~40% lower cost. The only time you need Opus is when the problem is genuinely novel.

Sources: Anthropic Official Blog, Anthropic System Card, SWE-bench, OSWorld, ARC-AGI-2, Vals AI, Artificial Analysis