Anthropic’s new Claude Opus 4.6 aims to think through bigger codebases

Anthropic is out with a new model called Claude Opus 4.6, an upgrade to its top-of-the-line Opus 4.5 model that launched in November. The new release could add new capabilities to Anthropic’s Claude Code coding assistant, which is facing growing competitive pressure from OpenAI’s Codex.

Anthropic says Opus 4.6 improves on its predecessor’s coding skills, planning, and, perhaps most importantly, its ability to reason more clearly when handling large amounts of information. When Opus 4.6 powers Claude Code, the coding agent can comprehend larger codebases and make more thoughtful decisions about how and where to add new code, the company says.

More long-term memory

AI labs have been racing to build models with longer context windows, meaning the amount of information a model can consider for a given task. But models have often struggled to use that information effectively in their outputs, a limitation Anthropic acknowledges.

“Previously, we would see things like, maybe the model gets lost in the middle, or it might forget details,” Opus product manager Dianne Penn tells Fast Company. “I wouldn’t say Opus 4.6 is perfect—humans or other past models aren’t perfect—but we think that the quality improvement is pretty significant.”

Opus’s longer memory also allows it to work on complicated tasks for extended periods, enabling Claude Code users to assemble teams of agents that collaborate on tasks. Anthropic also says the tool offers improved code review and debugging capabilities, helping it catch its own mistakes.

Opus 4.6 arrives as the use of AI coding tools continues to surge, and as competition between Anthropic and OpenAI for software developers intensifies. OpenAI’s Codex coding tool recently launched as a standalone app, powered by the GPT-5.2 model, and has received largely enthusiastic reviews from developers.

A model for everyday work tasks

Beyond coding, the new Anthropic model is designed to improve performance on everyday work tasks such as running financial analyses, conducting research, and creating or using documents, spreadsheets, and presentations. Opus 4.6 will also power Anthropic’s general-purpose work tool, CoWork, enabling it to multitask with minimal human supervision.

Anthropic says Opus 4.6 achieved top scores across several industry benchmark tests, reaching the highest results so far on multiple evaluations. These include Humanity’s Last Exam, a complex multidisciplinary reasoning test; Terminal-Bench 2.0, an agentic coding evaluation; and GDPval-AA, which measures performance on economically valuable knowledge-work tasks in finance, legal, and other domains. Anthropic also says Opus 4.6 outperforms all other models on OpenAI’s BrowseComp, which measures a model’s ability to locate difficult-to-find information online.

Anthropic says the Opus 4.6 model is available to developers using Claude Code for the same price per million tokens as Opus 4.5. The new model is now the default for Claude Code Pro subscribers, and is available as an option for all other subscribers. 

View the full article