What is Claude Sonnet 4.6, and what do free users get from it?

Yesterday, February 17, 2026, Anthropic released Claude Sonnet 4.6, two weeks after Opus 4.6. Sonnet is Anthropic's middle model: faster and cheaper than Opus, more capable than the small Haiku. TechCrunch notes the release keeps Sonnet on a roughly four-month update cycle.

The change most people will notice is simple: on the Free and Pro plans, Claude now answers with Sonnet 4.6 by default, on claude.ai and in Claude Cowork. If you use Claude without paying, you are now talking to this model.

What Anthropic released

Anthropic calls Sonnet 4.6 a full upgrade across coding, computer use, long-document reasoning, agent planning, office work and design. The essentials:

	Claude Sonnet 4.6
Who gets it	All Claude plans; default for Free and Pro
Price through the API	$3 per million input tokens, $15 per million output (unchanged)
Context window	1 million tokens, in beta
API name	`claude-sonnet-4-6`
Also in	Claude Code, Claude Cowork, major cloud platforms

The free plan also gains features that used to require more: creating files, connectors (links to outside apps and data), skills (packaged instructions for specific tasks) and compaction, which summarizes older parts of a long chat so the conversation can continue.

In early testing inside Claude Code, Anthropic's coding tool, testers picked Sonnet 4.6 over Sonnet 4.5 in about 70% of comparisons. More surprising, they preferred it to Opus 4.5, Anthropic's flagship from November, 59% of the time. They described it as less prone to overengineering (building more than was asked) and to "laziness," better at following instructions, and less likely to claim success falsely or to make things up.

Computer use, explained

Computer use means the AI operates a computer the way you do: it looks at screenshots and moves a virtual mouse and keyboard. That matters because so much software, from old billing systems to internal tools, has no API, no programmatic doorway another program can use. A model that can simply use the screen can work with all of it.

Anthropic first offered this in October 2024 and admitted it was experimental and error-prone. The standard test is OSWorld, hundreds of tasks in real programs such as Chrome, LibreOffice and VS Code on a simulated computer. Anthropic says its Sonnet models have improved steadily on it over sixteen months, and early users of Sonnet 4.6 report human-level results on tasks like navigating a complicated spreadsheet or completing a multi-step web form across several browser tabs. Anthropic is careful to add that it still trails the most skilled human users.

The risk is prompt injection: hidden instructions on a web page that try to hijack the AI. Anthropic says Sonnet 4.6 resists these much better than Sonnet 4.5, at a level similar to Opus 4.6.

Long context and long plans

The 1-million-token window can hold a whole codebase, a long contract or dozens of research papers at once. Anthropic's more interesting claim is that the model reasons well across all of it, which helps with long-range planning.

Its example is Vending-Bench Arena, a simulation in which AI models each run a small business over many months and compete for profit. Sonnet 4.6 found its own strategy: it spent heavily on capacity for the first ten simulated months, far more than its rivals, then switched hard to profitability near the end, and finished well ahead. A model that can delay rewards and change course at the right moment is better suited to long, open-ended tasks.

TechCrunch highlights another result: 60.4% on ARC-AGI-2, a test built to measure the kind of flexible reasoning that comes naturally to people. That beats most comparable models but remains below Opus 4.6, Google’s Gemini 3 Deep Think and a refined GPT-5.2 configuration.

Other updates

For developers: Sonnet 4.6 supports adaptive and extended thinking. On the API, web search and fetch tools now write and run code to filter results, keeping only what is relevant. Code execution, memory, tool search and programmatic tool calling are now generally available.
Claude in Excel now connects through MCP connectors to financial data services such as S&P Global, LSEG, PitchBook, Moody's and FactSet, on Pro, Max, Team and Enterprise plans.

Anthropic's safety researchers describe Sonnet 4.6's character as broadly warm, honest and prosocial, with very strong safety behavior and no major signs of high-stakes misalignment.

Try it: If you use Claude's free plan, give Sonnet 4.6 a task that used to need several messages, such as "Create a one-page budget spreadsheet for a family of four, then explain the three biggest categories." Download the file it creates and check whether the formulas work.

Should you use Claude Sonnet 4.6 or Claude Opus 4.6?

Start with Sonnet 4.6. Anthropic itself says it delivers performance that used to require an Opus model, including on real office tasks, at a lower API price: $3 and $15 per million input and output tokens, against $5 and $25 for Opus 4.6. For writing, research, everyday coding, spreadsheets and computer-use tasks, it is the sensible default, and on the free plan it is the model you already have.

Switch to Opus 4.6 when the task demands the deepest reasoning. Anthropic names three examples: refactoring a large codebase, coordinating several agents in one workflow, and problems where getting it exactly right matters more than speed or cost.

A simple test helps you decide for your own work:

Prompt

Here is a task I do every week: [describe it]. Do it once, then
list any assumptions you made and anything you were unsure about.

Run it with both models. If Sonnet's answer and list of doubts are as good as Opus's, stay with Sonnet. If Opus catches problems Sonnet missed, keep Opus for that kind of task.