Anthropic unveiled a new artificial intelligence (AI) model, Claude Opus 5, touting its "bang for the buck."
On the 24th (local time), Anthropic said on its official website that it is releasing Opus 5, offering performance close to the top public model, Fable 5, at about half the expense. Opus 5 outperformed existing public models on major benchmarks, including coding and knowledge work.
On Frontier-Bench v0.1, which evaluates an agent's terminal-environment coding ability, it scored 43.3%, beating OpenAI's GPT-5.6 Sol (34.4%) and Fable 5 (33.7%). On GDPval-AA v2, which evaluates knowledge-work ability, it scored 1,861 points, and it also earned the top mark in agent search performance with 90.8%.
However, on DeepSWE v1.1, which evaluates agent coding ability, it scored 68.8%, slightly lower than GPT-5.6 Sol (72.7%). On the "Hardest Last Exam (HLE)," which evaluates expert-level reasoning, it scored 56.3% without using external tools, narrowly trailing Fable 5 (56.5%).
Anthropic highlighted price as the model's biggest strength. For developers, Opus 5's API fees are set at $5 per 1 million input tokens and $25 per 1 million output tokens, the same as the transfer model Opus 4.8.
Anthropic's move to release this kind of "bang-for-the-buck" model is seen as a response to intensifying competition with high-performance, low-expense AI models from OpenAI, Google, xAI, and China's Kimi K3 and DeepSeek.
Meanwhile, Anthropic said Opus 5's ability to detect cybersecurity vulnerabilities is close to its internal research model, Mythos 5. However, it noted that its ability to turn detected vulnerabilities into real attacks and its long-duration autonomous research ability in biology are lower than Mythos 5.