Introduction

Anthropic recently launched the upgraded Claude model, Opus 4.7, intentionally positioning its capabilities below the highly anticipated Claude Mythos.

Opus 4.7 is marketed as a “significant enhancement” over Opus 4.6, with improvements in advanced software engineering, visual processing, memory, instruction following, and financial analysis.

However, the yet-to-be-officially-released Mythos has overshadowed Opus 4.7 to some extent. Notably, Anthropic has also kept a low profile regarding Opus 4.7, stating that it is “not as advanced” as the “Claude Mythos Preview” and has “more limited overall capabilities.”

This upgrade follows the release of the Glasswing project, Anthropic’s safety initiative that utilizes Claude Mythos Preview to identify and fix cybersecurity vulnerabilities.

Tech analyst Carmi Levy remarked, “It is rare in tech history for a product launch to focus on what it ‘cannot do’ rather than what it ‘can do.’ Anthropic’s message clearly indicates that Opus 4.7 is a safer model, with its capabilities intentionally limited compared to Mythos.”

Capability Enhancements and Safety Improvements

Anthropic emphasizes that Opus 4.7 has made “significant improvements” in instruction following compared to Opus 4.6, enabling it to handle complex, long-running tasks while maintaining “precise attention” to instructions. User feedback indicates they can delegate “the trickiest programming tasks” to this model, which also exhibits superior memory capabilities over its predecessor. The model can retain relevant notes across multi-turn conversations and apply them to new tasks, reducing the need for extensive context input.

Anthropic claims that Opus 4.7’s visual capabilities are three times better than those of previous models, supporting high-resolution image inputs of up to 2576 pixels, making it suitable for multimodal tasks requiring fine visual details, such as analyzing dense screenshots or extracting data from complex charts.

Additionally, the company states that Opus 4.7 excels in financial analysis, generating “rigorous analyses and models” and producing more professional-quality reports.

In terms of safety, Anthropic reports that Opus 4.7 is on par with its predecessor regarding the occurrence rates of undesirable behaviors such as “deception, flattery, and collusion abuse.” However, the company also notes that despite improvements in honesty and resistance to malicious prompt injection attacks, Opus 4.7 is “slightly weaker” than Opus 4.6 in other areas, such as responding to harmful prompts, and its behavior is still not fully ideal.

Comparison with Mythos

The release of Opus 4.7 coincides with high industry anticipation for Claude Mythos. Mythos is described as a general-purpose cutting-edge model, which Anthropic claims is the “best-aligned” model trained to date. Interestingly, in today’s release blog, Anthropic revealed that Mythos Preview outperformed Opus 4.7 in several key benchmark tests, with some differences exceeding ten percentage points.

Mythos Preview scored higher in SWE-Bench Pro and SWE-Bench Verified (agent programming), Humanity’s Last Exam (multidisciplinary reasoning), and BrowseComp (agent search); both models scored similarly in agent computer usage, graduate-level reasoning, and visual reasoning.

Opus 4.7 is now available across all Claude products and APIs, and it is also supported by Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. The pricing remains consistent with Opus 4.6: $5 per million input tokens and $25 per million output tokens.

Industry Positioning and Market Analysis

Yaz Palanichamy, a senior consultant analyst at Info-Tech Research Group, stated that Claude Opus is positioned in the industry as a “practical cutting-edge” model, representing Anthropic’s “most capable intelligent multifunctional automation model,” with core application scenarios covering complex programming, deep research, and integrated intelligent workflows.

He explained that the model’s core product differentiation advantage lies in its embedded algorithms’ coordination and composability when scaling various operational scenarios.

Palanichamy also noted that Claude Opus 4.7 is a “technology-oriented” platform requiring a certain degree of deep personalization to optimize prompts and generated outputs. Although Google Gemini 3.1 Pro has a larger context window (2 million tokens compared to Claude’s 1 million tokens), Opus 4.7 still maintains a clear advantage over Gemini in application engineering scenarios. He added that “some similar models are indeed converging in raw reasoning capabilities.”

Levy pointed out that the update to version 4.7 pushes Opus beyond basic chatbot workflows, positioning it closer to a “complex technical role co-pilot” tool. “It is more powerful than ever and serves as a better assistant for knowledge workers,” he said. At the same time, it presents lower risks, making it a “carefully balanced compromise.”

He also noted that the release of Opus 4.7 comes just two months after the launch of Opus 4.6, which itself “is a signal of how overheated the AI development cycle has become and how fierce market competition has become.”

Glasswing Security Project and Mythos Strategic Significance

Last week, Anthropic also announced the Glasswing project, applying Mythos Preview to defensive security. The company is collaborating with AWS, Google, and over 30 cybersecurity agencies to advance this project, claiming that Glasswing has identified “thousands” of high-risk vulnerabilities across every major operating system and web browser.

Anthropic intends to limit the release scope of Claude Mythos Preview, prioritizing testing new cybersecurity defense mechanisms on “weaker models,” with Opus 4.7 being one of them, as its cybersecurity capabilities are inferior to Mythos. In fact, Anthropic admits that it deliberately conducted experiments during training to “differentiate and reduce” these capabilities.

Anthropic explained that Opus 4.7 is equipped with protective mechanisms that can automatically detect and block requests involving “prohibited or high-risk” cybersecurity uses, with relevant experiences being applied to the Mythos model.

Levy believes this “acknowledges to some extent that the new model’s capabilities are intentionally reduced compared to its high-end counterparts” to enhance the credibility of its network risk detection and interception capabilities.

From a marketing perspective, this allows Anthropic to position Opus 4.7 as an ideal balance between capability and risk, without having to bear the “cybersecurity burden” that comes with the limited release of high-end models.

Levy stated that Mythos is likely to become the “ultimate groundwork” for the widespread adoption of Opus 4.7. Even if Mythos “is increasingly likely” never to be publicly released, it will serve as “an ideal means” to demonstrate that Opus is the model providing the best balance for most enterprise decision-makers.

Palanichamy agrees, noting that Opus 4.7 can serve as a public “testbed” for real-time testing and optimization of automated cybersecurity defense mechanisms, which will ultimately become a necessary prerequisite for the broader release of Mythos-level cutting-edge models.

Q&A

Q1: What are the main improvements of Opus 4.7 compared to Opus 4.6?
A: Opus 4.7 has achieved significant improvements in several areas: enhanced instruction following ability, capable of handling complex long-term tasks; visual capabilities three times better than its predecessor, supporting high-resolution images; stronger memory capabilities, able to remember notes across multi-turn conversations and apply them to new tasks; and improved financial analysis capabilities, generating more professional analysis reports.

Q2: What are the differences between Opus 4.7 and Claude Mythos?
A: Mythos Preview outperformed Opus 4.7 in several benchmark tests, with some differences exceeding ten percentage points, especially in agent programming (SWE-Bench), multidisciplinary reasoning (Humanity’s Last Exam), and agent search. Anthropic intentionally set Opus 4.7’s capabilities lower than Mythos, particularly reducing cybersecurity-related capabilities to ensure safer deployment.

Q3: On which platforms is Opus 4.7 available, and what is the pricing?
A: Opus 4.7 is now available across all Claude products and APIs, and it is also supported by Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. The pricing remains consistent with the previous Opus 4.6: $5 per million input tokens and $25 per million output tokens.