Claude Sonnet 4.6: 5 things it does better than the last version

Is it worth switching from Claude Sonnet 4.5 to the new 4.6? If you use the model for work, yes, and at no extra cost. Anthropic shipped Sonnet 4.6 on February 17 and kept pricing untouched, while adding features that solve real day-to-day problems.

Sonnet is the middle tier in Anthropic’s lineup, the one most people actually use. Opus is more expensive and slower, Haiku is faster but less precise. Sonnet sits right in the middle, which is why every new version matters for a regular user on the 20-dollar-a-month subscription.

1. One million tokens of context, beta option

This is the biggest change. The previous version had 200,000 tokens, enough for one longer document. With one million tokens, the model can ingest an entire novel or a mid-sized codebase in a single prompt. In practice, that means you can ask about something on page 600 of a document without summarising it first.

The feature is currently in beta, which at Anthropic usually means it works reliably but is not auto-rolled out to everyone.

2. Better results on SWE-Bench

SWE-Bench measures how well a model resolves real bug reports in actual software projects. Sonnet 4.6 set a new record on that test. Development teams with early access report 4.6 is more reliable than 4.5 on typical tasks, even compared with the pricier Opus 4.5 model from last November.

A concrete example from practice: writing a database migration script. The old version often needed several iterations. The new one gets it right on the first try more often.

3. Computer use

This is a feature Anthropic first showed off at the end of 2024, but it only becomes usable for real tasks with 4.6. The model can control keyboard and mouse, fill in web forms, switch between browser tabs. The score on the OS World benchmark, which measures exactly those tasks, is now a record.

In practical terms, you can tell the model to fill out twenty fields in a spreadsheet from another document and have it done without your intervention.

4. Better instruction following

This is the least flashy but most important change for daily work. The model is better at understanding when you said “short” or when you asked for a bullet list instead of prose. The old Sonnet 4.5 had a habit of drifting into essayistic tone when you did not ask for it. The new version sticks to the format.

If you draft emails through the model, this change is the one you feel most. Less editing after generation.

5. Price stays the same

Three dollars for a million input tokens, fifteen dollars for a million output tokens. Same numbers as the previous version. That is a rare move in the AI industry, where vendors usually charge more for a better model.

For a regular user on the ChatGPT-competitor app at claude.ai, Sonnet 4.6 is now the default model on Free and Pro plans. You do not have to pick anything.

What it means in practice

If you use Claude for work, the upgrade is automatic and free. If you are choosing between Claude and ChatGPT, this update puts Anthropic a step ahead again on software development and large-document work. If you are interested in agents that take action through your computer, the beta feature with the OS World record is the reason to try.

What developers should know

Anthropic accepts US payment cards for the 20-dollar subscription through claude.ai, with the same functionality as elsewhere. There are no geographic restrictions on content. An API key for development teams comes directly through console.anthropic.com.

The only practical wrinkle: subscriptions cannot be paid through the Apple App Store in some markets (downloading the Mac app is free, but the subscription has to go through the website), which is more of a minor detour than a real barrier.

The next version, according to unofficial chatter coming out of Anthropic, is expected in the third quarter. Until then, 4.6 is realistically the strongest mid-tier model on the market.