Anthropic Enhances Opus for Developers

Anthropic has introduced Claude Opus 4.7, an updated version of its widely accessible artificial intelligence model, positioning it as a more effective tool for advanced software engineering. This release follows the company's decision to restrict access to its more powerful model, Mythos Preview, due to concerns regarding cyber risks. The new version, Opus 4.7, is said to offer improvements over its predecessor, Opus 4.6, particularly in handling complex and lengthy coding tasks. Enhancements include better adherence to instructions, increased consistency, and improved self-checking mechanisms prior to delivering responses. The model became available on Thursday across various platforms, including Claude products, APIs, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry, with pricing remaining at $5 per million input tokens and $25 per million output tokens. This release strategy is noteworthy, especially considering the recent limitations placed on Mythos Preview, which Anthropic has characterized as a significant advancement in cyber capabilities. Under the initiative known as Project Glasswing, Mythos Preview is being treated as a restricted frontier system, following internal assessments that indicated its potential to surpass even highly skilled human experts in identifying and exploiting software vulnerabilities. By advancing Opus while keeping Mythos under tight controls, Anthropic appears to be aiming for broader adoption without exposing its most sensitive technologies to the public. The timing of this launch is influenced by customer demand, as Anthropic has increasingly marketed Claude as an enterprise-grade assistant for coding, research, and knowledge work. Opus 4.7 is presented as a practical enhancement for these applications rather than a significant public unveiling of its most powerful model. The company has emphasized endorsements from various software and enterprise users, highlighting improvements in handling complex engineering tasks, reducing tool errors, and enhancing reliability during extended sessions. The heightened emphasis on coding capabilities reflects a broader trend within the AI industry, where developers are competing not only on conversational fluency but also on becoming the preferred systems for software teams and corporate agents capable of executing multi-step tasks with minimal oversight. Anthropic asserts that Opus 4.7 can manage more challenging coding assignments that previously required closer supervision, introducing a new “xhigh” effort setting that allows users to prioritize deeper reasoning over speed for complex tasks. However, the company cautioned that the model's enhanced instruction-following may necessitate adjustments to older prompts, as it interprets commands more literally than earlier iterations. Safety considerations remain a key aspect of the model's development, although with some qualifications. Anthropic reports that Opus 4.7 maintains a safety profile similar to that of Opus 4.6, with improvements in honesty and resistance to prompt-injection attacks, while acknowledging some areas of weaker performance, such as providing overly detailed harm-reduction advice regarding controlled substances. Furthermore, it noted that internal evaluations suggest Mythos Preview may still exhibit better alignment, a significant acknowledgment given that this model is not yet available to the public. Concerns regarding Mythos have already reached financial and policy discussions. Reports indicate that U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell recently met with bank executives to address cyber risks associated with Anthropic's latest model. Additionally, access to Mythos is expected to be limited to approximately 40 technology companies, highlighting the unusual restrictions surrounding this major AI launch and the seriousness with which officials are treating the potential for AI-assisted cyber exploitation.

---