Anthropic Opus 5 released: performance close to Fable 5 but half the price, refreshing ARC-AGI-3 record

IT House reported on July 25 that Anthropic today released the latest flagship model of the Opus series - Claude Opus 5. Compared with the previous generation Opus 4.8, it has improved performance while maintaining the same price. Anthropic said that the Opus 5 is positioned as a high-performance model for daily use. Officials say that it is close to the higher-spec Claude Fable 5 in some capabilities, but the price is about half of the latter. This model has now become the default model for the Claude Max service and is the highest-capacity model available to Claude Pro users, reaching the current highest level in multiple code, knowledge work and scientific research tests. According to reports, Opus 5 has achieved leading results in code and knowledge work evaluations such as Frontier-Bench and GDPval-AA. In Frontier-Bench v0.1 testing, Opus 5 outperformed other models and more than doubled task performance compared to Opus 4.8 while costing less per task. In the CursorBench 3.2 test, the Opus 5's score in highest effort mode was just 0.5% lower than the Fable 5's peak, but at about half the cost per task. Additionally, the Opus 5 outperformed other models in performance-per-cost at high, medium-high, and maximum effort settings. In terms of knowledge processing and problem-solving capabilities, the test results published by Anthropic show that Opus 5 achieved high results in ARC-AGI 3, Zapier AutomationBench, OSWorld 2.0 and other evaluations. Among them, in the ARC-AGI 3 test, Opus 5's score is about 3 times that of the second place; in Zapier AutomationBench, which measures the model's ability to complete a complete business process, its pass rate is about the same as the cost.