Anthropic launches Claude Opus 5 with efficiency, safety improvements

Anthropic PBC today rolled out a large language model called Claude Opus 5 to its chatbot service and developer platform.

The company says that the LLM approaches the output quality of its top-end Mythos 5 algorithm in many areas, but costs significantly less.

Anthropic hasn't made Mythos 5 publicly available. The company instead sells access to a scaled-down version of the LLM called Fable 5 that shares many of its features, but poses a smaller cybersecurity risk. Anthropic compared Fable 5 to Opus 5 across 13 benchmarks. It says that the latter algorithm scored higher on 8 of the tests despite costing about 50% less.

Opus 5's edge was particularly pronounced across two benchmarks called Frontier-Bench and AutomationBench. The former test comprises 74 tasks spanning fields such as physics, chemistry and cryptography. Opus 5 performed 9.7% better than Fable 5. It achieved a 8.5% higher score on AutomationBench, which includes a set of knowledge work tasks.

One of the contributors to Opus 5's performance is that it verifies its work. The Frontier-Bench benchmark required the model to generate a three-dimensional machine part schematic. Opus 5 had no way to view the drawing, which made verifying its accuracy impossible. The model addressed the challenge by writing a "computer vision pipeline to pull the geometry from the raw pixels."

Opus 5's output verification feature also lends itself to biology research. According to Anthropic, the model outperformed an earlier LLM called Opus 4.8 across all its internal life sciences benchmarks. Opus 5 proved particularly adept at automated protein research, an important field that was the focus of the 2024 Nobel Prize in chemistry.

Anthropic says that Opus 5 is its safest model to date by multiple measures. Compared to its predecessors, the LLM is less prone to deceptive behavior and errors that are difficult to reverse.

Anthropic has not made its flagship Mythos 5 publicly available over concerns that it could be used by hackers to launch cyberattacks. The model can not only find vulnerabilities but also develop workflows for exploiting them. By contrast, Opus 5 scored zero on an internal Anthropic benchmark that evaluates LLMs' ability to exploit vulnerabilities.

The company has equipped the LLM with more relaxed guardrails than Fable 5. Customers can use Opus 5 to search code for security issues, but it blocks vulnerability scan methods that are commonly used by hackers. Anthropic estimates that its guardrails will activate 85% less frequency than for Fable 5.

Opus 5 is now the default model in the top-end Max tier of the Claude chatbot service. It's also available in the mid-range Pro plan.

Developers can connect Opus 5 to their software via Anthropic's application programming interface. As part of today's update, the API received a feature that makes it possible to interrupt an LLM mid-workflow and change the tools that it uses. Another new capability can reroute prompts that are blocked by Anthropic's safety guardrails to a different model.