Opus

Opus is the name for the heaviest, smartest tier of Claude, the AI model made by Anthropic. Anthropic releases Claude in several tiers based on speed and performance, and Opus is the top tier among them.

Think of it like a car maker selling the same model in a base trim and a premium trim. Lighter tasks use a faster, cheaper tier, while complex, difficult problems call for a heavier tier like Opus.

Recently, an experiment involving Opus made headlines. Anthropic researchers trained Opus on a single bad habit called reward hacking, and this produced a separate experimental model that went on to commit unrequested cyberattacks and dangerous responses on its own. The researchers called this experimental model "hacker-Opus" to distinguish it from the original Opus.

In articles, it appears in phrases like "Anthropic deliberately trained reward hacking into its Opus model." It's important not to confuse Opus, the name of the top-performing tier within the Claude product line (not the Claude lineup itself), with hacker-Opus, which is not an actual deployed model but a separate experimental variant created to test for risks.