Meet Claude Sonnet 4.6: Smarter, Faster, and Built for the Real World

The Wait Is Over

On February 17, 2026, Anthropic quietly changed the game. Claude Sonnet 4.6 became the new default model on claude.ai — and if you've used it, you probably noticed immediately. Conversations feel crisper. Complex tasks finish faster. And for the first time, using a Sonnet-class model feels like reaching for the best tool in the shed, not a compromise.

So what exactly changed? Quite a lot, it turns out.

A Million Tokens of Memory

Let's start with the headline number: 1 million tokens. That's the size of Sonnet 4.6's context window, now available in beta. To put that in perspective, it's roughly equivalent to feeding the model a full-length novel, a 200-page legal document, and your entire codebase — all at once.

In real-world tests, the model handles this enormous context with remarkable accuracy. In MRCR v2 testing (a benchmark that hides information deep inside long documents), Sonnet 4.6 achieved a ~65% Mean Match Ratio — up from just 18.5% in its predecessor, Sonnet 4.5. That's not a small improvement. That's a transformation.

For enterprise users, this means compliance reviews, financial audits, and long-horizon coding tasks can now happen in a single conversation, without losing track of what came before.

Code Like Never Before

Developers, this one's for you.

Sonnet 4.6 scores approximately 79.6% on SWE-bench Verified — a rigorous benchmark of real-world software engineering tasks. That puts it in territory previously reserved for Opus-class models. Early access developers reported preferring Sonnet 4.6 over its predecessor by a wide margin, and many even preferred it over Anthropic's previous top-tier model, Claude Opus 4.5.

The improvements aren't just about raw scores. Consistency, instruction-following, and multi-step reasoning have all been sharpened. If you've ever had a model drift off-task halfway through a complex refactor, you'll appreciate the difference.

Computer Use: Finally, Actually Useful

Back in October 2024, Anthropic introduced the world's first general-purpose computer-using AI model. At the time, they called it "still experimental — at times cumbersome and error-prone."

Sonnet 4.6 is the payoff on that promise.

The model now scores 72.5% on OSWorld-Verified, a benchmark for autonomous computer use tasks. Users are reporting human-level reliability on things like navigating complex spreadsheets, filling out multi-step web forms, and orchestrating workflows across multiple browser tabs simultaneously.

This matters because most enterprise software was built before modern APIs existed. Sonnet 4.6 can operate those legacy systems the way a person would — no custom connectors required. It's also significantly more resistant to prompt injection attacks than Sonnet 4.5, making it safer to deploy in automated environments.

Speed That Changes the Workflow

Raw intelligence only matters if the model can keep up with you.

Sonnet 4.6 is 30–50% faster than Sonnet 4.5 on average, with independent benchmarks clocking latency as low as 0.98–1.41 seconds and throughput around 39–42 tokens per second. For comparison, Opus 4.6 runs at 1.8–2.6 seconds — meaning you get near-Opus intelligence at Sonnet speed.

For agentic workflows where a model may make dozens of sequential decisions, that latency difference compounds into a dramatically better experience.

Same Price, More Power

Perhaps the most remarkable part of the Sonnet 4.6 launch: the pricing didn't change. At $3 per million input tokens and $15 per million output tokens, it matches Sonnet 4.5 exactly.

You're getting a model that approaches Opus-level performance on many tasks — including real-world office work — at a fraction of the cost. That's the kind of efficiency gain that makes previously impractical deployments suddenly viable.

Safety First, Always

Every new Claude model goes through extensive safety evaluations before launch. Sonnet 4.6 is no exception. Anthropic's safety researchers described the model as having "a broadly warm, honest, prosocial, and at times funny character, very strong safety behaviors, and no signs of major concerns around high-stakes forms of misalignment."

Beyond character evaluations, Sonnet 4.6 showed major improvements in resisting prompt injection attacks compared to Sonnet 4.5 — a critical consideration as the model is increasingly deployed in autonomous, agentic settings.

The Bottom Line

Claude Sonnet 4.6 isn't a minor iteration. It's a rethinking of what a mid-tier AI model can do. It brings:

-> A 1 million token context window for handling massive documents and long conversations

->~80% SWE-bench performance for serious software engineering tasks

72.5% OSWorld scores for near-human computer use

-> 30–50% speed improvements over its predecessor

-> Unchanged pricing from Sonnet 4.5

For developers, enterprises, and everyday users who want the best performance without paying Opus-class prices, Sonnet 4.6 is the new default — in every sense of the word.

It's available now on claude.ai (Free, Pro, and Team plans), the Claude API, Amazon Bedrock, and Google Vertex AI.