Claude Opus 4.6 builds on earlier releases with improved coding performance and more consistent behavior in complex tasks.
Opus 4.6 finds real vulnerabilities in codebases better than any other model (Source: Anthropic)
According to Anthropic, the model applies more deliberate planning during task execution, sustains agent-driven workflows over longer periods, and operates with greater consistency across large codebases. It improves code review and debugging by identifying errors in its own output and correcting them during execution.
The model also applies these improvements to common professional tasks. It can run financial analyses, support research workflows, and create or work with documents, spreadsheets, and presentations. These functions benefit from stronger reasoning across extended inputs, allowing the model to process large volumes of material and continue working with that information over time.
“We build Claude with Claude. Our engineers write code with Claude Code every day, and every new model first gets tested on our own work. With Opus 4.6, we’ve found that the model brings more focus to the most challenging parts of a task without being told to, moves quickly through the more straightforward parts, handles ambiguous problems with better judgment, and stays productive over longer sessions,” the company said.
Advances in safety assessment
Safety evaluations remain a core part of the release. Automated behavioral audits show low rates of misaligned behavior, including deception, sycophancy, reinforcement of user delusions, and cooperation with misuse.
The model maintains alignment comparable to earlier releases and shows a lower rate of over-refusals, where benign queries go unanswered.
The release includes the most extensive safety testing applied to a Claude model to date. Evaluations cover user wellbeing, complex refusal behavior for potentially harmful requests, and assessments of covert harmful actions. New interpretability methods were also tested to better understand model behavior and identify issues that standard testing might miss.
Additional safeguards were applied in areas where the model’s capabilities could be used in both defensive and harmful ways.
Claude Opus 4.6 is also being used to support cyber defensive work, including identifying and patching vulnerabilities in open-source software.
“Early testing shows Claude Opus 4.6 delivering on the complex, multi-step coding work developers face every day, especially agentic workflows that demand planning and tool calling. This starts unlocking long-horizon tasks at the frontier,” said Mario Rodriguez, CPO, GitHub.
