Anthropic launches Claude Opus 4.7 as safer mid tier model

Anthropic has released Claude Opus 4.7, calling it the company's strongest generally available model with significant improvements in advanced software engineering and coding tasks. The new model, which became available Thursday across all Claude products, the company's API, and through cloud providers Microsoft, Google, and Amazon, outperforms its predecessor Opus 4.6 across multiple benchmarks, including agentic coding where it scored 64.3% on SWE-bench Pro and SWE-bench Verified tests. Despite these improvements, Anthropic acknowledges that Opus 4.7 is "less broadly capable" than Claude Mythos Preview, a restricted model currently available only to select organizations through a cybersecurity initiative called Project Glasswing that launched earlier this month. The release highlights Anthropic's strategic focus on software engineering capabilities, with users reporting they can now hand off their most difficult coding work with confidence. Opus 4.7 demonstrates improved performance in complex, long-running tasks, better instruction-following, and enhanced tool use and agentic computer use. The model also features substantially better vision capabilities and produces more creative outputs for tasks like slides, documents, and interfaces. Notably, Opus 4.7 includes safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses, which Anthropic says resulted in slightly lower performance on cybersecurity vulnerability reproduction benchmarks compared to Opus 4.6, dropping from 73.8% to 73.1%. The launch appears designed in part to showcase the capabilities of the restricted Mythos model, which outperforms Opus 4.7 and other competing models in nearly every benchmark category. Anthropic is encouraging security professionals interested in advanced cybersecurity capabilities to apply through a formal verification program. Opus 4.7 maintains the same pricing as Opus 4.6, though the company notes it uses an updated tokenizer that may result in higher token usage and potentially increased costs depending on usage patterns.