American artificial intelligence company Anthropic announced the launch of a new generation of large model Claude Opus 4.7 on Thursday, saying that it has improved in many capabilities over previous versions, but overall it is still "not as broadly powerful as its strongest product, Claude Mythos Preview." The new model focuses on software engineering, executing complex instructions and processing real business work scenarios. It is positioned as the company's current "strongest general model open to the public".
In contrast, Claude Mythos Preview focuses on network security and offensive and defensive capabilities, and is regarded by Anthropic as the most powerful cutting-edge model currently. It was only opened to some enterprise customers on a limited basis earlier this month through the "Project Glasswing (Glass Wing Plan)" network security project, which is used to identify software vulnerabilities and network attack risks. In order to reduce the risk of abuse, Anthropic made it clear that Claude Opus 4.7 deliberately weakened its "network security-related capabilities" during training, and its network attack and defense capabilities are not as good as the Mythos-level model.
Anthropic pointed out in the statement that the company has added an automated security protection mechanism to Claude Opus 4.7, which can detect and intercept requests involving prohibited or high-risk network security purposes. It hopes that by observing and evaluating the operation of these security mechanisms in real environments, it will gradually accumulate experience for the larger and more secure release of Mythos-level models in the future. Anthropic said that since its establishment in 2021, the company has been committed to releasing and iterating large models in a safe and controllable manner, and has deliberately distinguished itself from competitors such as OpenAI as a vendor that "pays more attention to security and responsible deployment."
Around the launch of Project Glasswing and the Mythos model, the White House, technology company executives and large bank CEOs have recently held multiple closed-door meetings to focus on the potential risks of cutting-edge AI models in cybersecurity and the financial system. Under the leadership of the current President Trump's administration, conversations around AI security among regulators, financial institutions and technology companies have become increasingly frequent, and Anthropic has become one of the important participants in the discussion with its security positioning.
The release of Claude Opus 4.7 isn't that long ago. Just this February, Anthropic launched Claude Opus 4.6. The company said that the new version surpasses 4.6 in a number of key scenarios, including industry-wide benchmarks such as agent coding, multi-disciplinary complex reasoning, large-scale tool invocation, and automated computer operations. In the process of training Opus 4.7, Anthropic made an attempt to "differentially weaken network security capabilities", trying to strengthen general task capabilities while selectively reducing its strength in high-risk network offense and defense.
Still, Anthropic doesn't plan to make Claude Mythos Preview fully available to the general public. The company has emphasized many times that its current goal is to continuously accumulate experience through small-scale pilots and explore paths for deploying Mythos-level models on a larger scale in the future. For security professionals who wish to use Claude Opus 4.7 for cybersecurity testing and defense within a legal compliance framework, Anthropic encourages application through a formal qualification program so that companies can evaluate use cases under a more stringent oversight environment.
At present, Claude Opus 4.7 has been launched in Anthropic's various Claude products. It can also be accessed through API and provided externally through cloud service providers such as Microsoft, Google and Amazon. In terms of pricing, Anthropic said that the new model will continue to use the price of Claude Opus 4.6 without any additional price increases, hoping to lower the threshold for enterprises and developers to upgrade to the new version.