Anthropic Unveils Claude Opus 4.7: Enhanced Vision and Security

Anthropic Innovates with Claude Opus 4.7

Anthropic continues to stand out in the field of artificial intelligence with the launch of Claude Opus 4.7. This model, now available on Claude, the API, and major cloud platforms, marks a significant advancement in vision and safety. Two months after the releases of Opus 4.6 and Sonnet 4.6, and a week after the presentation of Claude Mythos Preview, deemed too dangerous for public release, Anthropic offers a model that promises to transform interaction with visual data.

Enhanced Vision and Autonomy

One of the major improvements in Claude Opus 4.7 lies in its ability to process high-resolution images, up to 2,576 pixels on the long side, or about 3.75 megapixels. This evolution allows the model to accept images over three times the resolution of previous models. It is particularly beneficial for AI agents that need to analyze dense screenshots or extract data from complex diagrams. Furthermore, the model excels in handling complex and prolonged tasks, following instructions with increased accuracy and verifying its results before submission. It is also more creative and meticulous in producing interfaces, presentations, and documents.

Instruction Tracking and Necessary Adjustments

Instruction tracking has been sufficiently improved that developers will need to readjust their prompts. Prompts written for earlier models may now yield unexpected results, as the model interprets them more literally. Anthropic's benchmarks show that Opus 4.7 outperforms Opus 4.6 across all tests, although it still falls short of Claude Mythos Preview. Pricing remains unchanged: $5 per million input tokens and $25 for output. The model is accessible via Claude, the API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry.

Token Consumption and Migration

Anthropic warns of a potential increase in token consumption when migrating from Opus 4.6. The new tokenizer may require up to 1.35 times more tokens for the same input text. The model also reasons more deeply, thereby increasing the volume of output tokens. Although internal tests indicate a net positive effect, Anthropic recommends monitoring the impact on actual traffic. A migration guide is available to facilitate this transition.

Cybersecurity and Project Glasswing

In terms of cybersecurity, Claude Opus 4.7 incorporates safeguards inherited from Project Glasswing, introduced last week. This initiative aims to deploy Mythos Preview to around fifty partner organizations, such as AWS, Apple, Microsoft, CrowdStrike, and the Linux Foundation, to secure critical software infrastructures. Anthropic has experimented with efforts to specifically reduce the model's cyber capabilities compared to Mythos Preview. The lessons learned from this deployment are intended to guide the eventual public availability of models in the Mythos range. The model is designed to detect and block risky queries while allowing security professionals to participate in the Cyber Verification Program for legitimate uses such as vulnerability research, penetration testing, and red teaming.

New Features and Tools

Anthropic also introduces several new features with Opus 4.7:

• An “xhigh” effort level for more precise reasoning control via the API.
• The /ultrareview command in Claude Code to identify bugs and design issues.
• Task budgets in public beta on the API to manage token consumption on long tasks.
• The extension of auto mode to Max subscribers in Claude Code for uninterrupted execution of long tasks.