Now loading...
Claude Opus 5 has officially launched today, offering a sophisticated model that approaches the advanced intelligence level of Claude Fable 5 at a fraction of the cost. This model is regarded as the new state-of-the-art for tasks in coding and knowledge work, although it still falls short in cybersecurity compared to Mythos 5. Designed for daily use, Opus 5 shows enhanced efficiency compared to its predecessors, serving as the default model on Claude Max and as the top model on Claude Pro.
Performance-wise, Claude Opus 5 significantly boosts productivity at the same price point as the earlier version, Opus 4.8. Users can optimize performance through customizable effort settings, balancing for either intelligence or cost-efficiency. Remarkably, Opus 5 outperforms existing models in specific software engineering tasks. For instance, it excels on the Frontier-Bench v0.1, exceeding Opus 4.8’s performance by more than double while reducing task costs. Moreover, on CursorBench 3.2, it operates at nearly the same efficiency as Fable 5 but at 50 percent of the cost per task.
The evaluation results highlight Opus 5’s prowess in knowledge work and problem-solving situations. Notably, it scores three times higher than its nearest competitor on ARC-AGI 3, and on Zapier AutomationBench, it outpaces the next best model by 1.5 times for similar costs, even in lower-effort settings. Additionally, on OSWorld 2.0, Opus 5 consistently achieved superior results compared to all other models at lower costs, surpassing Fable 5 by a significant margin.
Improvements extend to various life sciences evaluations, particularly regarding organic chemistry and protein-related tasks, indicating Opus 5’s enhanced capability for scientific research. Moreover, it produces remarkably strong visual outputs.
When engaging with Claude Opus 5, users have observed its enhanced ability to verify its own work and thoroughly iterate until achieving successful outcomes. In trials, Opus 5 demonstrated agency and meticulousness, such as developing its own computer vision pipeline to reconstruct complex machine parts without direct visual access, and successfully identifying and addressing code issues in an open-source package manager.
Feedback from early access users has highlighted Opus 5’s capabilities. It reportedly approaches Fable-level performance at a lower cost on specific benchmarks, such as FrontierCode 1.1, and excels in challenging tasks like debugging. Industry analysts have noted the model’s stronger analytical capabilities, with improvements in tasks that require critical thinking and organizational skills.
Regarding alignment and safety, Opus 5 is recognized as highly aligned per its automated behavioral audit, showing reduced misalignment compared to prior models. Although it does exhibit advanced capabilities, it remains less adept in handling risks related to dual-use technology, particularly in cybersecurity and biological research, where it still trails behind Mythos 5.
Claude Opus 5 is available today across all platforms, with a pricing structure identical to Opus 4.8, at $5 per million input tokens and $25 per million output tokens. It can run in a Fast mode, which increases processing speed at double the base price. Additionally, two beta updates are being rolled out: mid-conversation tool changes and automatic request fallbacks for safety. Comprehensive guidance for maximizing Opus 5 is accessible via the official prompting guide.
