SAN FRANCISCO — Anthropic, the AI safety-focused startup, unveiled its latest artificial intelligence model, Claude Opus 4.7, on Thursday, positioning it as a significant upgrade for everyday applications while deliberately dialing back its potential for advanced cybersecurity tasks compared to a more powerful but restricted predecessor.
The company described Claude Opus 4.7 as its "most powerful generally available model" to date, excelling in areas like software engineering, instruction-following, and real-world task completion. According to Anthropic's official release, the model outperforms its immediate forerunner, Claude Opus 4.6, which launched in February, across key benchmarks including agentic coding, multidisciplinary reasoning, scaled tool use, and agentic computer use. This release marks another step in Anthropic's iterative approach to building AI systems that balance capability with caution, a hallmark of the firm since its founding in 2021.
However, Anthropic was upfront about the model's limitations. In a statement, the company noted that Claude Opus 4.7 is "broadly less capable" than Claude Mythos Preview, a cutting-edge model rolled out earlier this month to a select group of companies under the banner of Project Glasswing, Anthropic's new cybersecurity initiative. Mythos Preview, according to reports from CNBC, excels particularly at identifying weaknesses and security flaws within software, making it a potent tool for defensive cybersecurity work—but one that raises concerns about misuse in offensive scenarios.
"We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses," Anthropic said in its announcement. The firm emphasized that insights from deploying these safeguards in real-world settings will inform future efforts toward a broader rollout of Mythos-class models. For now, Mythos Preview remains unavailable to the general public, with Anthropic stating its goal is to learn from controlled deployments before scaling up.
This cautious rollout comes amid heightened scrutiny over AI's role in national security. The launch of Project Glasswing has prompted a series of high-profile discussions involving members of the Trump administration, tech executives, and banking leaders, focusing on the risks posed by increasingly powerful AI systems. Sources familiar with the matter, as reported by CNBC, indicate these meetings underscore growing worries about AI's potential to exacerbate cyber threats, especially as models like Mythos demonstrate advanced capabilities in probing software vulnerabilities.
Anthropic's approach reflects its long-standing commitment to responsible AI development, setting it apart from competitors like OpenAI. Founded by former OpenAI executives including siblings Dario and Daniela Amodei, the company has prioritized safety research from the outset, investing heavily in techniques to align AI with human values. This philosophy has earned Anthropic partnerships with major cloud providers and a reputation for measured progress, even as rivals race ahead with more ambitious releases.
During training, Anthropic experimented with methods to "differentially reduce" Claude Opus 4.7's cyber capabilities, ensuring it doesn't match Mythos's prowess in sensitive areas. The company is encouraging security professionals interested in leveraging the model for "legitimate cybersecurity purposes" to apply through a formal verification program. This initiative aims to channel AI's power toward defensive applications, such as vulnerability detection, while mitigating risks of it being used for hacking or other illicit activities.
Claude Opus 4.7 is now accessible across Anthropic's suite of Claude products, its application programming interface (API), and through integrations with cloud giants Microsoft, Google, and Amazon Web Services. Pricing remains unchanged from the Opus 4.6 model, making it an attractive option for developers and businesses seeking reliable AI assistance without the premium costs associated with experimental previews like Mythos.
The timing of this release is notable, arriving just weeks after Project Glasswing's debut and amid a broader conversation about AI governance. Earlier this month, Anthropic co-founder Dario Amodei spoke at a Semafor event, indicating the company is "in process of broadening Mythos access" to trusted partners. This gradual expansion aligns with Anthropic's strategy of testing safeguards in limited environments before wider deployment, a practice that has drawn both praise and criticism from industry watchers.
Critics argue that such self-imposed restrictions could hinder innovation, especially as global competition in AI intensifies. OpenAI, for instance, has faced its own controversies over safety lapses, including incidents where models generated harmful content. In contrast, Anthropic's transparency about Opus 4.7's reduced capabilities—coupled with built-in blocks on risky queries—positions it as a leader in ethical AI, though some experts question whether these measures are sufficient against sophisticated users.
From a technical standpoint, the improvements in Opus 4.7 are substantial. Benchmarks show gains in handling complex, multi-step tasks, which could benefit sectors like software development and data analysis. For example, the model's enhanced agentic coding abilities allow it to autonomously write, debug, and optimize code more effectively than previous versions, according to Anthropic's evaluations.
Yet, the shadow of Mythos looms large. Described internally as Anthropic's most advanced offering, the preview model was designed specifically for cybersecurity under Project Glasswing, which involves collaboration with select enterprises to simulate real-world defensive scenarios. By limiting Opus 4.7's similar functions, Anthropic aims to prevent unintended escalations in cyber warfare, a concern echoed in recent White House briefings on AI risks.
Looking ahead, Anthropic has no immediate plans for general availability of Mythos Preview, but the company views Opus 4.7 as a bridge to that future. "What we learn from the real-world deployment of these safeguards will help us work towards our eventual goal of a broad release of Mythos-class models," the release stated. This iterative process could shape industry standards, influencing how other firms approach AI safety in high-stakes domains like cybersecurity.
The announcement also highlights ongoing tensions in the AI landscape. While Anthropic's safeguards are proactive, questions persist about enforcement. Security experts, speaking on condition of anonymity, have noted that no system is foolproof, and determined actors might find workarounds. Nonetheless, Anthropic's verification program for professionals signals a structured path forward, potentially fostering greater trust among regulators and users.
As AI continues to permeate critical infrastructure—from banking systems to national defenses—the release of Claude Opus 4.7 underscores the delicate balance between innovation and restraint. With Trump administration officials actively engaging on these issues, future policy could mandate similar safeguards, reshaping the competitive dynamics of the $100 billion-plus AI market. For now, Anthropic's measured step forward offers a glimpse of what's possible when safety takes precedence, even if it means holding back the full potential of models like Mythos.