Claude Sonnet 4.6 Drops: Faster Output, Same Opus Intelligence

Anthropic released Claude Sonnet 4.6 on February 17, twelve days after Opus 4.6 dropped. The positioning is familiar: Sonnet is the faster, cheaper tier that handles the majority of real-world tasks without requiring the full weight of Opus. With Sonnet 4.6, Anthropic is making a stronger claim than usual that the intelligence gap between the two tiers has narrowed.

Speed Is the Main Story

Sonnet 4.6 generates output significantly faster than Opus 4.6. In internal testing across standard coding and writing tasks, output speed runs roughly 2 to 2.5 times faster than Opus 4.6. For interactive use like writing code, answering questions, or drafting documents, that difference is noticeable.

In agentic workflows where a model might run dozens or hundreds of turns, the speed difference compounds. A task that takes Opus 4.6 ten minutes might take Sonnet 4.6 four or five. For most production applications, that’s the relevant comparison.

Intelligence Parity Claims

Anthropic says Sonnet 4.6 matches Opus 4.6 on most tasks. That’s a strong statement, and it’s worth being specific about what it means.

On the MMLU benchmark (general knowledge), both models score above 90%. On coding benchmarks like HumanEval, Sonnet 4.6 scores within two to three percentage points of Opus 4.6. For everyday tasks like summarization, question answering, drafting, and code explanation, the outputs are often indistinguishable.

Where they differ is at the hard ceiling. On SWE-bench Verified, Opus 4.6 scores 80.8% against Sonnet 4.6’s 79.6%. That gap is narrow, only about 1.2 percentage points, but it still favors Opus for autonomous software engineering agents where you’re trying to maximize the fraction of issues that get resolved correctly. For a developer using the model interactively (writing code with it, reading its output, iterating) the difference is effectively invisible, because human judgment fills in the gap.

Reasoning-heavy tasks show the biggest gaps between the two tiers. Deep multi-step mathematical proofs, complex logic problems, and very long-document analysis all tend to favor Opus 4.6. For most software development and content work, Sonnet 4.6 gets the job done.

Pricing Comparison

At launch, Sonnet 4.6 cost $3 per million input tokens and $15 per million output tokens. That’s about 40% cheaper than Opus 4.6 on both input ($5/M) and output ($25/M).

For a production application making 10,000 API calls per day at average context sizes, the cost difference is substantial. Sonnet 4.6 makes Claude viable for higher-volume use cases that Opus pricing would rule out.

The correct mental model: use Sonnet 4.6 as the default for most things, route to Opus 4.6 for tasks where quality ceiling matters most.

Context Window

Sonnet 4.6 ships with the full 1 million token context window, matching Opus 4.6. That’s a notable change from earlier Sonnet generations, which were capped well below the flagship tier. You can fit hundreds of pages of documents, entire codebases, or extensive conversation histories without hitting a wall.

The 1M token window is most valuable for specific workloads: full-repository analysis, very long document processing, or extended agentic sessions that accumulate large amounts of context. With Sonnet 4.6 you get that headroom at Sonnet pricing, which removes one of the historical reasons to reach for Opus.

Claude Code Integration

At release, Claude Sonnet 4.6 became the default model for Claude Code, Anthropic’s terminal-based coding agent. The previous default was Sonnet 4.5, so this was a capability upgrade for Claude Code users who hadn’t manually selected a different model. (Anthropic has since shipped newer flagships, including Opus 4.8 in May 2026, so the in-product defaults have moved on.)

I’ve been running Sonnet 4.6 through Claude Code for the past week on typical development tasks: debugging, refactoring, writing tests, documentation. The response quality is noticeably better than 4.5 for complex multi-file operations. The speed improvement over Opus 4.6 makes iteration feel faster, which matters when you’re in a rapid back-and-forth debugging session.

Where Things Stand

The Sonnet/Opus pattern has been Anthropic’s strategy for a few model generations now, and with each generation the Sonnet tier has gotten closer to Opus on raw capability. Sonnet 4.6 continued that trend right on schedule. At its launch it was the right default for most developers, faster, cheaper, and genuinely capable across the full range of typical coding and writing tasks. Anthropic has since released newer models, including its direct successor Claude Sonnet 5 on June 30, 2026 and the current flagship Claude Opus 5 on July 24, so check the models overview for the latest tier before choosing.

The 80.8% vs 79.6% SWE-bench gap is real but narrow, and if you’re building autonomous coding agents where every point of issue resolution rate counts, Opus 4.6 still earns its price premium. For everything else, Sonnet 4.6 is the practical choice.