Opus 4.7: A Leap Forward in AI Capabilities
Anthropic is thrilled to announce the general availability of Opus 4.7, our latest AI model. This release represents a significant advancement over Opus 4.6, particularly in complex software engineering tasks. Users are reporting increased confidence in delegating their most challenging coding projects to Opus 4.7, previously requiring close supervision.
Enhanced Performance and Reliability
Opus 4.7 demonstrates remarkable rigor and consistency when handling intricate, long-running tasks. It exhibits a heightened attention to detail in following instructions and proactively implements methods to verify its outputs before delivering results. This improved reliability makes it a powerful tool for developers and professionals alike.
Superior Vision and Creative Output
Beyond coding, Opus 4.7 boasts substantially improved vision capabilities, processing images with greater resolution. It also excels in creative applications, producing higher-quality interfaces, presentations, and documentation. The model is more refined and tasteful when completing professional tasks, delivering polished and impactful results.
Cybersecurity Safeguards and the Cyber Verification Program
While Opus 4.7 isn’t as broadly capable as our most powerful model, Claude Mythos Preview, it consistently outperforms Opus 4.6 across various benchmarks. Recognizing the potential risks associated with advanced AI in cybersecurity, we’ve prioritized safety. Opus 4.7 is the first model to incorporate safeguards designed to automatically detect and block requests related to prohibited or high-risk cybersecurity activities. This is part of our phased approach, testing these safeguards on less capable models before a broader release of Mythos-class models.
We’ve announced Project Glasswing, outlining the benefits and risks of AI in cybersecurity. Security professionals seeking to utilize Opus 4.7 for legitimate purposes – such as vulnerability research, penetration testing, and red-teaming – are encouraged to join our new Cyber Verification Program.
Availability and Pricing
Opus 4.7 is available today across all Claude products, and through our API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. Pricing remains consistent with Opus 4.6: $5 per million input tokens and $25 per million output tokens. Developers can access Opus 4.7 via the Claude API.
Early Tester Feedback & Safety Profile
Early access testers have provided overwhelmingly positive feedback on Opus 4.7. Our evaluations indicate a similar safety profile to Opus 4.6, with low rates of concerning behaviors like deception and susceptibility to malicious prompt injection. In some areas, such as honesty, Opus 4.7 demonstrates improvement. A comprehensive overview of our safety evaluations can be found in the Claude Opus 4.7 System Card.
Important Updates: Token Usage
While Opus 4.7 is a direct upgrade to Opus 4.6, two changes impact token usage:
- Updated Tokenizer: The new tokenizer improves text processing, but the same input may now map to more tokens (approximately 1.0–1.35x depending on content type).
- Increased Effort: Opus 4.7 operates at higher effort levels, particularly in agentic settings, enhancing reliability but potentially increasing output token usage.
Users can manage token usage through the effort parameter, task budgets, or by prompting for conciseness. Our internal testing shows overall token usage is improved, but we recommend monitoring performance with your own traffic. A detailed migration guide is available to assist with upgrading from Opus 4.6.
1 This is a model-level change, so images sent to Claude will be processed at higher fidelity automatically. To minimize token consumption, users can downsample images before sending them if high resolution isn’t required.
