What are Claude Opus 4 and Claude Sonnet 4, and what can they do?

Yesterday, Thursday, May 22, 2025, Anthropic introduced two new models in the fourth generation of Claude, Claude Opus 4 and Claude Sonnet 4. Anthropic presents them as a step toward a "virtual collaborator," an AI that can stay on a long project, keep track of what it has done and carry real work forward rather than answer one question at a time.

Both models are on claude.ai, with Sonnet 4 open to free users, and both are available to developers through Anthropic's own API as well as Google Cloud's Vertex AI and Amazon Bedrock. Prices are unchanged from the previous Opus and Sonnet models.

What Anthropic released

The two models share a design but serve different needs:

The scores are Anthropic's own. SWE-bench Verified asks a model to fix real bugs from open-source projects and checks the fix with tests. Sonnet 4 edges Opus 4 on this one test, but Anthropic says Opus 4 is stronger in most areas and on long tasks. A token is a small piece of text, roughly part of a word.

Anthropic also announced several companions to the models:

• Claude Code is now generally available, no longer a preview. New beta extensions show its edits directly inside the VS Code and JetBrains code editors, and a software kit lets developers build their own agents on the same engine.
• New developer tools in the API: a tool that runs code, a connector for MCP (the Model Context Protocol, a standard way to plug AI into outside tools and data), a Files API, and longer prompt caching, which saves money when the same long instructions are sent repeatedly.

How the models work

Both are hybrid models. For a simple request they reply almost at once. For a harder one you can switch on extended thinking, in which the model reasons step by step before answering.

The new part is that thinking and acting can now be interleaved. In a beta feature, Claude can pause its reasoning to run a web search or another tool, read the results and continue thinking, much like a researcher who stops mid-thought to check a reference. It can also call several tools at the same time.

Memory is the second advance. When a developer gives Claude access to files on a computer, Opus 4 can write and update its own notes, such as key facts and progress so far, and consult them later. Anthropic's example is Opus 4 keeping a navigation guide while playing the video game Pokémon. For long jobs, this kind of notebook helps the model stay consistent over hours instead of minutes.

Anthropic also reports that both models are 65% less likely than Claude 3.7 Sonnet to take shortcuts or loopholes, such as making a test pass without truly solving the problem, on tasks prone to that behavior. And when Claude's thinking runs very long, a smaller model now summarizes it for display, which Anthropic says happens only about 5% of the time.

A worked example: a long task

Anthropic says Opus 4 can work continuously for several hours, and that Rakuten had it run an open-source code refactor, a large reorganization of existing code, on its own for 7 hours. CNBC's report describes that as nearly a full workday. A request that suits this kind of model looks less like a question and more like a work order:

Our app's date handling is spread across 40 files and uses two different
libraries. Plan a migration to one library. Keep a notes file with your
plan, what you have changed and what is left. After each group of files,
run the tests and fix any failures before moving on. Stop and ask me
before deleting anything you are unsure about.

The notes file uses the memory ability; running tests after each step lets the model check its own work; the final line keeps a person in charge of risky choices.

Why it matters

The release shows where Anthropic is putting its effort. Jared Kaplan, its chief science officer, told CNBC that Anthropic ended its investment in chatbots late last year to concentrate on complex work such as research and coding. He also acknowledged the risk: the longer and more complex the task, the more chances for a model to go off the rails.

Partners are already adopting the models. GitHub says Sonnet 4 will power the new coding agent in GitHub Copilot, and companies such as Cursor, Replit and Sourcegraph praised its coding.

Safety is part of the story. For the first time, Anthropic is deploying a model, Opus 4, under its stricter AI Safety Level 3 (ASL-3) protections. The company says this is a precaution: it has not concluded that Opus 4 needs them, but could no longer clearly rule it out, mainly over the risk of helping someone with chemical, biological, radiological or nuclear weapons. The measures include filters that watch for such requests and tighter security around the model's weights, the trained numbers that make up the model. Sonnet 4 stays under the standard ASL-2 level.

Should you use Claude Opus 4 or Claude Sonnet 4?

For most people, Sonnet 4 is the right starting point. It is fast, much cheaper through the API, available on the free plan, and Anthropic's own coding score puts it level with Opus 4 on everyday programming fixes. Writing, summarizing, analysis and ordinary coding all fit it well.

Choose Opus 4 when the task is long, complex or high-stakes: a large code change across many files, a research project with many sources, or an agent that must work for hours with little supervision. It costs five times as much per token, so it pays to save it for jobs where its extra stamina and depth matter.

Whichever you pick, try extended thinking for hard problems, give the model clear goals and limits, and review the results before you rely on them.