Claude Opus 4.7 - A New Frontier, in Performance … and Drama

Opus 4.7 shifts behavior via “adaptive thinking”: it spends less compute on tasks it deems easy. In practice, that changes output depth and can miss necessary work unless you explicitly ask for it. You also can’t force long reasoning; effort defaults are now “medium,” which quietly affects quality and latency tradeoffs.

Benchmarks are mixed and over-indexed by marketing. Opus 4.7 improves long-context reasoning in some tasks, but regresses on agentic browsing and certain retrieval tasks. It underperforms cheaper Gemini 3 Flash on OCR, and Anthropic intentionally reduced cybersecurity vulnerability reproduction. Takeaway: performance is now workload-specific; don’t assume a single “best” model.

Operationally, compute scarcity appears to be biting. Reports of throttling, availability variance, and forced adaptive reasoning suggest capacity management is influencing product behavior. Sudden deprecations (Opus 4.5, 4.0) create migration risk. Safety reviews are under time pressure; internal anecdotes highlight hallucination/fabrication and coordination risks in agentic settings—manage with scaffolds and governance.

For coding, Claude remains strong on real-world codebases and “last-mile” usability, but OpenAI claims it’s caught up by training on messy repos. Expect tighter competition, price/perf spreads by task, and more value in orchestration than any single model. Teams relying on browsing/vision should consider cheaper, better-per-task alternatives and set reliability SLOs.

Net: treat 4.7 as a specialized tool within a routed workflow, not a universal upgrade. Control effort settings, measure outcome quality (not output volume), and prepare for vendor-side compute constraints to shape behavior without notice.