Anthropic has released Claude Opus 4.7, an upgraded version of its flagship AI model with stronger coding and vision capabilities. The model can handle complex software engineering tasks more independently and now supports image inputs up to 2,576 pixels on the long edge. It also introduces built-in safeguards to detect and block high-risk cybersecurity use. While it improves on Opus 4.6, it remains less powerful than Anthropic’s restricted Mythos Preview model.
Building on the foundation laid by Opus 4.6, the new model shows clear gains in handling multi-step programming workflows that previously required close human supervision. These include tasks like large-scale code refactoring, debugging across multiple files, and maintaining logical consistency in complex systems over extended interactions.
Another major highlight of Opus 4.7 is its improved multimodal capability. By increasing the maximum image resolution it can process to 2,576 pixels on the long edge, over three times the limit of earlier Claude models, the system becomes far more effective in visually intensive tasks. This includes interpreting dense dashboards, analyzing UI screenshots, reading technical diagrams, and assisting with frontend development.
The Dario Amodei-led firm has introduced several technical upgrades to improve developer control and usability. A new ‘xhigh’ effort level sits between the existing high and max settings, allowing users to better balance reasoning depth with response speed. The company has also launched task budgets in public beta, allowing API users to limit how much compute a task can use, helping manage costs in long-running or agent-based workflows. Commercially, Anthropic has kept pricing unchanged from Opus 4.6, maintaining rates at $5 per million input tokens and $25 per million output tokens. Even Claude Code now includes an ‘ultrareview’ command for deeper code inspection and more effective bug detection.
In terms of performance, Opus 4.7 shows clear improvements over Opus 4.6, scoring higher in finance-focused agent evaluations and on GDPval-AA, a benchmark measuring AI performance in economically valuable tasks like finance and legal work. Another key change is a new tokenizer, which can produce between 1.0 and 1.35 times more tokens for the same input, depending on content type.
However, despite all such advances, Anthropic has been explicit about the model’s positioning within its broader AI portfolio. Opus 4.7 does not represent the company’s most powerful system. That distinction belongs to Mythos Preview, a more advanced model that remains under restricted access due to safety concerns.
Safety features are a key part of the latest release. The model includes systems that automatically detect and block prompts related to prohibited or high-risk cybersecurity activities. Compared to Mythos Preview, its cyber capabilities have been deliberately reduced during training to lower the risk of misuse. At the same time, Anthropic has introduced a Cyber Verification Program, allowing vetted security professionals to access the model for legitimate use cases like vulnerability testing and defensive research. The latest model is being released across Claude apps, its API, and major cloud platforms like Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.