Anthropic Unveils AI ‘Opus 5’...Surpasses Latest OpenAI Model

Rated Higher Than Fable by Analytics Firm
Pricing at Only Half That of Fable

AI company Anthropic has unveiled its new AI model, Claude Opus 5.

Yonhap News

On July 24 (local time), Anthropic announced the launch of its AI model, Opus 5, stating that the model achieves nearly the same performance as its flagship public model, Fable 5, but at just half the cost.

Opus 5 outperformed OpenAI's latest model, GPT-5.6 Sol, in various performance benchmarks across fields such as coding and knowledge work. In some benchmarks, it was even rated higher than Fable 5.

In the “Frontier-Bench v0.1” test, which measures coding capability in agent terminal environments, Opus 5 scored 43.3%, surpassing both GPT-5.6 Sol (34.4%) and Fable 5 (33.7%).

It also achieved a higher score than the other two models in "GDPval-AA v2," which evaluates knowledge work ability, with 1,861 points, and recorded the highest score of 90.8% for agent search performance.

However, in the "DeepSWE v1.1" indicator, which reflects agent coding ability in general environments, Opus 5 scored 68.8%, trailing GPT-5.6 Sol's 72.7%. In the “Final Exam for Humanity” (HLE, no tools), which measures expert-level reasoning skills, Opus 5 scored 56.3%, slightly below Fable 5’s 56.5%.

Its performance in areas of cybersecurity, which have become a recent focus of concern, has also improved significantly.

Anthropic explained, “Even though we deliberately avoided training Opus 5 on cybersecurity tasks, improvements in its overall capabilities have led to advancements in these (security) areas as well.”

However, the company emphasized that while Opus 5’s ability to discover security vulnerabilities is at a similar level to “Mythos 5”—the top-tier model available only to select trusted partners—its hacking skills, or the ability to exploit these vulnerabilities for cyber threats, are significantly lower than those of Mythos 5.

It also noted that important limitations remain when it comes to extended, autonomous scientific research that could potentially contribute to the development of biological weapons and related threats.

AI performance analysis firm Artificial Analysis assigned Opus 5 an intelligence score of 61, one point higher than Fable 5, ranking it first among publicly available AI models.

The developer pricing for Opus 5 remains at $5–25 per 1 million tokens, the same as its predecessor Opus 4.8, which is half the price of Fable 5. This is also comparable to or slightly lower than competitor OpenAI’s GPT-5.6 Sol, which charges $5–30 per 1 million tokens.

Anthropic is releasing the Claude AI models in a detailed performance-based hierarchy, ranging from the top-tier Mythos and Fable models, through Opus and Sonnet, down to the lightweight Haiku model.

This content was produced with the assistance of AI translation services.