Anthropic released Opus 5 on July 24, and the company's own technical notes, echoed by Ars Technica's writeup, frame it as a token-efficiency release, not a jump in what the model can do. That distinction matters because the two things cost differently: a capability jump means the model solves harder problems it couldn't before, while an efficiency gain means it solves the same problems using fewer output tokens per task, which is a direct line item on every enterprise API bill. Simon Willison's technical breakdown the same day backs the efficiency read: benchmark scores move modestly, but cost-per-completed-task drops. The Verge's coverage adds one data point worth naming precisely: Opus 5 is described as nearing Fable 5 capabilities, Anthropic's other model line, which tells you the gap Anthropic is closing is internal, between its own two products, not against OpenAI or Google's newest release.

For enterprise deployment teams, the relevant number is dollars per resolved ticket or per completed coding task, not a leaderboard rank. A bank or a systems integrator running Claude in production reprices its inference budget on July 24 pricing, not on whatever benchmark chart accompanies the launch post. The open item is whether Anthropic publishes a per-token price cut alongside the efficiency claim, or leaves customers to compute the savings themselves from the smaller token counts. Watch Anthropic's pricing page for a Q3 2026 update tied to this release; that page, not the benchmark table, is where the Opus 5 story actually settles for procurement.