AI long-distance runner: Anthropic presents Claude 4

Contents

A year is a long time in the context of AI development. While other developers have been throwing new models around in the meantime, Anthropic has remained remarkably faithful to its third generation of the Claude Opus model. Now the company, which emerged from former OpenAI employees in 2021, has launched the next major version leap with Sonnet 4 and Opus 4.

Opus 4 is said to be – who is surprised – the most powerful AI system from Anthropic to date. What is more exciting is that, according to Chief Product Officer Mike Krieger, the difference should be particularly noticeable for existing customers.

Opus 4: Seven hours of continuous running

Of course, Anthropic cannot escape the “agentic age” recently proclaimed by Microsoft and Google. Both of Anthropic's new models are designed to perform tasks more autonomously. Anthropic has given its Opus model more time so that it can also handle more complex tasks. It can implement plans lasting several hours without suffering an interruption. In sample projects, the model worked on a programming project for around seven hours. Users should also be able to choose between quick answers and more time-consuming, human-like thought processes.

According to media reports, Anthropic wants to remain true to its strategy of scoring points in the turbulent AI market with the outstanding capabilities of its models and a focus on programming aids. Behind the scenes, however, the company struggled for a while to make decisive progress with its strongest Opus model. This is why updates have been released for the other models meanwhile, but not for Opus.

Ahead in programming tests

The Claude 4 models have been leading in real software engineering tasks in tests, Anthropic advertises the new generation. Opus 4 scored 72.5 percent of correct solutions in the SWE-Bench, ahead of other models. GPT-4.1 from OpenAI (69.1 percent) and Gemini 2.5 Pro (63.2 percent) were significantly behind in some cases. Anthropic presents further test results in a blog post.

With their introduction, the possibility of using tools such as web search in extended thought processes is also being tested. Both models can also use tools in parallel, should follow instructions more precisely and — if developers are given access to local files — have significantly improved storage capacities. Users of the Sonnet model should also notice significant improvements after the jump from version 3.7 to 4.

Anthropic has also made its coding assistant Claude Code generally available. The tool now supports background tasks via GitHub Actions and native integrations with VS Code and JetBrains.

API prices remain constant

For developers using the Anthropic API, some new capabilities have been unlocked, such as the MCP connector or the Files API. Fortunately, Anthropic is keeping its previous prices from the previous generation: Opus 4 costs $15 or $75 per million tokens (input/output) to use, while Sonnet 4 costs $3 or $15 per million tokens (input/output).

Sonnet is available to users free of charge via the web. Opus is only available as a paid subscription.

(mki)