Anthropic Releases Claude Opus 4.7 — Powerful Code Model with Reduced Cybersecurity Capabilities

Anthropic has released Claude Opus 4.7 — the company's most powerful AI model available on the mass market, though not the most powerful in principle. In parallel, the company has effectively acknowledged that its more advanced Mythos Preview model is currently too risky for mass rollout due to its strong cyber capabilities.

Therefore, the new release appears to be a compromise between practical power and safety: Opus 4.7 is tailored for complex development, long agent tasks, and enterprise work, but comes with additional restrictions on dangerous cyber scenarios. The release took place on April 16, 2026, just nine days after the announcement of Mythos Preview on April 7, 2026.

Mythos is what Anthropic describes as its most powerful and simultaneously most security-sensitive AI. Instead of broad access, the company kept Mythos in the limited circle of Project Glasswing, where the model is tested by selected partners working with mission-critical software and infrastructure. The logic is simple: first give defenders a head start, and only then decide whether such a system can be safely scaled to the mass market.

At the product level, Opus 4.7 is a direct upgrade from Opus 4.6 with an emphasis on software engineering.

Anthropic says the model handles the most difficult programming tasks better, requires less oversight, follows instructions more precisely, and can maintain complex multi-step context longer. There is a separate emphasis on the fact that the model more frequently verifies its own result before returning the answer. For visual tasks, there is also an upgrade: Opus 4.

7 accepts images with side length up to 2576 pixels, or roughly 3.75 megapixels, which is more than three times higher than the limit for previous Claude models. This is important for agent scenarios where models need to read dense screenshots, diagrams, and interfaces.

Anthropic also added a new xhigh mode between high and max, allowing developers to fine-tune the balance between reasoning depth and response latency. The model is already available in Claude, through the API, and also in Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The price remains the same relative to Opus 4.

6: 5 dollars per million input tokens and 25 dollars per million output tokens. For corporate teams, this is an important signal: the company did not simply roll out another demo update, but provided a compatible commercial upgrade without an immediate tariff revision. According to Anthropic's own tests, Opus 4.

7 provides approximately 13% improvement on their internal benchmark of 93 programming tasks relative to Opus 4.6, including four tasks that Opus 4.6 and Sonnet 4.

6 previously did not solve. But the most interesting point is not in the performance figures, but in the fact that the company openly emphasizes the gap with Mythos. Anthropic directly states that Opus 4.

7 is less universal and less capable in the cyber domain than Mythos Preview, and during training the team separately experimented with reducing precisely such capabilities. On top of this, the model received automatic protective mechanisms that should recognize and block requests related to prohibited or high-risk use in cybersecurity. For legitimate researchers, the company has a separate verification program.

This is an important shift for the entire market. Not long ago, AI developers competed mainly by showing the best benchmark and rolling out their flagship to public access faster. Now a different logic is emerging: the most powerful model does not necessarily become mainstream if, along with the growth in quality, its usefulness to attackers grows too quickly.

In Anthropic's case, a two-tier scheme results. Mythos remains in limited access as a tool for protected and controlled scenarios, while Opus 4.7 becomes the public workhorse for code, automation, and enterprise agents.

For developers, this is good news because they get a noticeably more powerful tool right now. For the industry as a whole — a reminder that the next race in AI will go not only for intelligence, but also for how carefully and in measured doses companies can release their most powerful systems.