AIBRIEF
Anthropic's new Claude line model ships as the practical default: nearly as capable as its top model at about half the cost, with safety classifiers needing to step in about 85% less often.
Anthropic released Opus 5 on Friday, positioning it as the new default in its Claude model family rather than a flagship replacement. The company says the model comes within a margin of its most capable system and prices it at roughly half the cost (Anthropic announcement; Engadget).
Anthropic says classifiers intervene on Opus 5 about 85% less often than on its top model, and the company expects Opus 5 to become the model "most people reach for" precisely because it removes that friction. On capability, Anthropic claims Opus 5 outperforms its predecessors on knowledge work and complex coding, and frames it as a stronger tool for scientific tasks like predicting how protein sequence variations affect function (Anthropic announcement).
That positioning coexists with two other Anthropic claims. The company says Opus 5 "exhibits the lowest rates of deceptive behavior" of any model it has shipped and was deliberately not trained on cyber-offensive tasks, even though it is broadly better at finding vulnerabilities. The more restricted "Mythos"-class model above it is currently limited to vetted organizations through Anthropic's Project Glasswing initiative (Engadget).
Anthropic's Claude platform pricing page lists current API rates; BenchLM tracks third-party prices. The "half the cost" framing is Anthropic's own claim, not yet an independently verified comparison.