Claude 4: Tests, Features, Access, Benchmarks, and More

Anthropic just released Claude 4, which comes in two versions: Claude 4 Sonnet and Claude 4 Opus.

Claude Sonnet 4 is a generalist model that’s great for most AI use cases and especially strong at coding. I think it’s one of the best models you can use for free.

Claude Opus 4 is designed for reasoning-heavy tasks like agentic search and long-running code workflows. Anthropic calls Opus 4 “the best coding model in the world,” but I find that claim a bit empty.

Yes, it’s currently the top performer on the SWE-bench Verified benchmark. But with a context window of just 200K, I can’t imagine it handling very large codebases cleanly. And let’s be honest: there’s always another stronger model coming out every month or so. Claiming the crown for a few weeks doesn’t make much sense.

That said, Claude 4 is still a very strong release. I’ll walk you through the most important details—features, use cases, benchmarks—and I’ll also run a few tests of my own.

We keep our readers updated on the latest in AI by sending out The Median, our free Friday newsletter that breaks down the week’s key stories. Subscribe and stay sharp in just a few minutes a week:

Claude Sonnet 4

Claude Sonnet 4 is the smaller model in the Claude 4 family. It’s designed for general-purpose use and performs well across most common AI tasks—coding, writing, question answering, data analysis. It’s also available to free users, which makes it unusually accessible for a model of this quality.

The model supports a 200K context window, which allows it to handle large prompts and maintain continuity over long interactions. That’s useful for use cases like analyzing long documents, reviewing codebases, or generating multi-part responses with consistent structure. However, Sonnet 4 might struggle with large codebases. For comparison, Gemini 2.5 Flash has a context window of 1M tokens.

Compared to Claude Sonnet 3.7, this version is faster, better at following instructions, and more reliable in code-heavy workflows. It supports up to 64K output tokens, which helps with slightly longer outputs like structured plans, multi-part answers, or large code completions.

Early reports show fewer navigation errors and better performance across app development tasks. It’s not as strong as Opus 4 when it comes to complex reasoning or long-term task planning, but for most workflows, it’s more than sufficient.

Claude Opus 4

Claude Opus 4 is the flagship model in the Claude 4 series. It’s built for tasks that require deeper reasoning, long-term memory, and more structured outputs—things like agentic search, large-scale code refactoring, multi-step problem solving, and extended research workflows.

Like Sonnet 4, it supports a 200K context window, so this can be a disadvantage if you want to use it with a large codebase. For comparison, Gemini 2.5 Pro (Google’s flagship model) has a context window of 1M tokens.

It’s also capable of running in “extended thinking” mode, where it switches from fast responses to slower, more deliberate reasoning. This mode allows it to perform tool use, track memory across steps, and generate summaries of its own thought process when needed.

Anthropic has positioned it as a high-end model for developers, researchers, and teams building AI agents. It leads on SWE-bench Verified and Terminal-bench, and early users report strong performance in coding agents, search workflows, and multi-hour tasks like refactoring open-source projects or simulating long-term planning.

Unlike Sonnet, Opus 4 is only available on paid plans. It’s more expensive to run, and likely overkill for simple chatbot use. But for applications that need consistent reasoning across a lot of moving parts, it’s the more capable choice.

Testing Claude 4

When testing a new model, I typically use the same tasks—that way I can see how it compares to other models I’ve tested before. This isn’t meant to be an extensive evaluation. The goal is just to get a feel for how these models behave inside the chat interface.

Let’s try Claude 4 on two areas: math and coding.

Math

I like to start with a simple calculation that often confuses language models. This isn’t about checking basic arithmetic—I could just use a calculator for that. The point is to see how the model approaches a slightly tricky problem, and whether it can fall back on tool use or show its reasoning clearly when needed.

Let’s see how well Claude Sonnet 4 handled it:

As you can see, it got the answer wrong on the first attempt. But when I asked it to use a tool—a calculator—it responded by writing a one-line script in JavaScript and solved the problem correctly.

Claude Opus 4 answered correctly on the first attempt.

Next, I wanted to see how well Claude Sonnet 4 handles a more complex problem: use all digits from 0 to 9 exactly once to make three numbers x, y, z such that x + y = z.

After about five minutes of random brute-force attempts, I got a message saying the output limit had been reached and I needed to click “Continue” to resume. I did, and Claude tried again—but then hit the limit once more. What I appreciated, though, is that it didn’t make up an answer. It simply refused to answer if it couldn’t find one. That’s a big win, in my opinion—it’s more problematic to hallucinate a solution.

Then I tried Claude Opus 4 on the same task. The answer came back almost instantly, and it was correct: 246 + 789 = 1035. Opus 4 is impressive!!

Coding

For the coding task, I decided to go straight to Claude Opus 4. This kind of creative generation feels more suited to its capabilities. I’m not testing it on large codebases here—just a relatively trivial coding task.

I asked it to make a quick p5.js game using this prompt that I used for Gemini 2.5 Pro and o4-mini:

Prompt: Make me a captivating endless runner game. Key instructions on the screen. p5.js scene, no HTML. I like pixelated dinosaurs and interesting backgrounds.

Normally, I’d copy-paste the code into an online p5.js editor to test it. But one of the nice features in Claude 4 is Artifacts, which lets me view and run the code output directly inside the chat.

Let’s see the result: