Thought Leadership
Anthropic has released Claude Opus 5, a new AI model designed to deliver performance close to the company’s more expensive flagship model, Claude Fable 5, at half the price.
According to Anthropic, Opus 5 sets new benchmarks across multiple evaluations. In software development tests such as Frontier-Bench v0.1, the model reportedly outperforms all competitors while doubling the score achieved by Opus 4.8, all at a lower cost per task. On CursorBench 3.2, Opus 5 in maximum-effort mode comes close to Fable 5’s performance at half the price per task.
Opus 5 also performs strongly on general problem-solving tasks. In the ARC-AGI-3 test, it reportedly scores three times higher than the next-best model. On Zapier AutomationBench, which measures a model’s ability to complete business tasks autonomously, Anthropic says Opus 5 achieves roughly 1.5 times the success rate of its competitors at the same price. For computer-use tasks measured by OSWorld 2.0, the model reportedly even surpasses Fable 5’s top score at just over one-third of the cost.
The model also brings improvements in life sciences, including deriving molecular structures from spectroscopy data and predicting protein functions.
Anthropic is also showcasing a video in which Opus 5 visualizes airflow around aerodynamic and less aerodynamic objects inside a virtual wind tunnel. The demonstration is intended to highlight the model’s improved visual capabilities.
Safety and Alignment
According to Anthropic, Opus 5 is the company’s best-aligned model to date. Automated behavioral audits reportedly found the lowest levels of deceptive behavior and susceptibility to misuse among all of the company’s models so far.
In security-related capabilities, Opus 5 deliberately remains behind the separate Mythos 5 model. While internal tests show that Opus 5 can now identify software vulnerabilities almost as effectively as Mythos 5, it falls significantly behind when it comes to actually developing exploits. Anthropic attributes this to a deliberate decision not to train the model specifically for cyberattack techniques.
Anthropic says the safety classifiers for Opus 5 intervene around 85 percent less frequently than those for Fable 5. Permitted use cases include identifying vulnerabilities in source code, while binary-based scanning, penetration testing, and exploit creation remain blocked. Blocked requests in Claude.ai, Claude Code, and Claude Cowork automatically fall back to Opus 4.8 by default.
How Much Does Claude Opus 5 Cost?
Claude Opus 5 is now available across all platforms. Pricing is the same as for Opus 4.8: $5 per million input tokens and $25 per million output tokens. Through the Claude API, the model is available under the name claude-opus-5. A “Fast” mode offering roughly 2.5 times the standard speed costs twice the base price.
Anthropic is also announcing two beta features: the ability to switch available tools in the middle of a conversation without losing the prompt cache, and automatic API fallbacks that route requests blocked by safety classifiers to another model automatically.
(ll)
