Anthropic has unveiled its latest large language model (LLM), Claude Opus 4.5, positioning it as a top-tier competitor in coding, agentic tasks, and general computer use. The release comes amid rapid advancements in the field, following recent updates from Google (Gemini 3) and OpenAI. While early claims suggest Opus 4.5 surpasses competitors in certain coding benchmarks, real-world performance data remains limited as it has not yet been extensively evaluated on platforms like LMArena.
Enhanced Capabilities & New Tools
The new model showcases significant improvements in deep research, slide-based workflows, and spreadsheet manipulation. Anthropic is simultaneously launching updates to Claude Code, its specialized coding tool, and its consumer applications, enabling more robust long-running agents and expanded functionality within tools like Excel, Chrome, and desktop environments. Opus 4.5 is now accessible via Anthropic’s platforms, APIs, and major cloud providers.
Cybersecurity Remains a Critical Challenge
Like all agentic AI tools, Claude Opus 4.5 continues to grapple with inherent cybersecurity vulnerabilities, particularly prompt injection attacks. These attacks exploit LLMs by embedding malicious instructions within external data sources, potentially overriding safety protocols and causing harmful actions, such as unauthorized data disclosure. Anthropic asserts that Opus 4.5 is more resistant to these exploits than other leading models, but acknowledges it is not immune.
Safety Evaluation Results: A Mixed Picture
Anthropic conducted internal and external safety evaluations to assess the model’s resistance to malicious prompts. In agentic coding scenarios, Opus 4.5 successfully refused 100% of 150 prohibited requests (e.g., generating harmful code). However, performance in more applied contexts was less consistent:
- Claude Code: Refused approximately 78% of prompts involving malware creation, DDoS attacks, and non-consensual surveillance software.
- Computer Use: Refused roughly 88% of prompts requesting unethical actions like surveillance, data collection, and generating harmful content (including simulated extortion attempts).
These results highlight a persistent gap between theoretical safety measures and real-world exploitation risks. Even with improved defenses, LLMs remain susceptible to manipulation, raising questions about the long-term viability of AI agents in sensitive environments.
The launch of Claude Opus 4.5 underscores the accelerating pace of AI development, but also reinforces the need for continuous vigilance regarding cybersecurity and ethical deployment. While Anthropic’s model represents a step forward in capabilities, vulnerabilities remain that must be addressed to ensure safe and responsible use.
