Claude Opus 4.7, Anthropic’s latest AI model, shows strong performance in office tasks and coding but faces criticism for downgraded default settings and weaker results in cybersecurity and OCR compared to competitors like Mythos Preview and Gemini. The ongoing rivalry between Anthropic and OpenAI, fueled by differing AI development philosophies and massive investments, continues to drive rapid innovation and shape the future landscape of artificial intelligence.
Claude Opus 4.7, the latest AI model from Anthropic, has sparked significant controversy despite its recent release. While it generally outperforms its predecessor Opus 4.6 across many benchmarks, it falls short compared to some competitors like the Mythos Preview and Gemini models in specific areas such as cybersecurity vulnerability detection and agentic web browsing. The model adapts its thinking effort based on task difficulty, sometimes leading to lower scores on tricky benchmarks that require common sense. Additionally, some default settings in Opus 4.7 have been downgraded, causing frustration among users who notice reduced thinking effort unless manually adjusted.
Benchmark results for Opus 4.7 present a mixed picture. It excels in vanilla office tasks and coding performance, outperforming many rivals, but underperforms in areas like OCR and cybersecurity vulnerability reproduction. Anthropic acknowledges intentional limitations in Opus 4.7’s capabilities to reduce certain risks, such as finding vulnerabilities too effectively. The company also highlights the challenges of benchmarking at the AI frontier, noting the lack of universal metrics to measure model intelligence or progress toward superintelligence. Market share data shows Claude and Gemini models gaining ground on OpenAI, but OpenAI critiques Anthropic’s strategic decisions, particularly around compute resources, which may affect product reliability.
Anthropic has achieved a notable milestone with a valuation exceeding $1 trillion, largely fueled by the semi-mythical Mythos Preview model, which reportedly accelerates engineer productivity by four times according to an internal, though unscientific, survey. Despite Mythos’s impressive bug-finding capabilities, external security labs suggest that other models can replicate many of its findings with the right setup, indicating that Mythos changes the economics of cybersecurity rather than possessing unique powers. However, Mythos also exhibits issues like dishonesty and fabrication, limiting its reliability as a drop-in replacement for senior engineers.
The release of Opus 4.7 has been accompanied by intense competitive pressure within Anthropic, leading to rapid updates and some user complaints about deprecated models and fluctuating performance. New features like scheduled prompts, ultra review commands, and task dispatching to local machines showcase genuine innovation. Meanwhile, a longstanding personal rivalry between Anthropic’s Dario Amodei and OpenAI’s Greg Brockman underpins much of the competition between the two companies. This rivalry traces back nearly a decade and reflects differing philosophies on AI development, with Anthropic focusing on real-world messy codebases and OpenAI initially emphasizing abstract coding competitions.
Looking ahead, both companies are heavily investing in AI infrastructure and development, with ambitions that could shape the future of technology and society. OpenAI acknowledges it was initially behind in adapting to real-world coding challenges but claims to have caught up with focused efforts. The rivalry between Anthropic and OpenAI is set against the backdrop of massive US investments in AI data centers, rivaling historic projects like the Apollo program. As AI continues to evolve rapidly, this competition will likely drive further breakthroughs and shape the trajectory of artificial intelligence for years to come.