Amid growing concerns from enterprise customers about expensive AI bills, Opus 5 comes with a feature enabling users to toggle how much effort—low, medium, or high—the model expends completing a task or answering a prompt. The company said this will enables users to balance between cost and capability.

Anthropic also said the model was easier to use than its predecessors, requiring less back and forth. It verifies its own work and recovers from errors without intervention.

This is the fourth model Anthropic has released in less than two months, following Mythos 5, Fable 5, and Sonnet 5 in June.

Fable 5 is the controversial, powerful model the U.S. government temporarily imposed export controls on after Amazon researchers reported they could bypass its safeguards. In response, Anthropic took the model off the market on June 12, and then re-released it on June 30 after fortifying its security.

Business customers and developers also criticized Fable 5’s burn rate, or the high number of tokens the model tended to use when completing tasks. This resulted in users blowing through their allotted token budgets or running up large bills. (Tokens are the units of information an AI model processes, equivalent in English text to about a word and a half. Tokens are also the unit AI companies tend to use to bill for model usage.)

Opus 5 “comes close” to the capabilities of Fable 5, but at half the price, according to Anthropic. However, the company still recommends Fable 5 for more advanced projects, including those which the model may handle autonomously for days on end.

OpenAI also emphasized economical token usage when marketing its most recent model, GPT-5.6, released on July 9.

On safety, Anthropic is also toeing a line between being safe, but not being too cautious, to the point of refusing legitimate safety work. Some users had criticized Anthropic for setting the guardrails around Fable 5 too stringently, making the model less useful for many tasks, particularly around coding and cybersecurity, some biological research tasks, and AI model development itself.

As with Fable 5, when Opus 5 declines a user’s request due to a safety concern, the API automatically falls back to another model, so the user gets an answer rather than an error.

But Anthropic said Opus 5’s safeguards with similar to those of Opus 4.8 in most ways, except in the area of cyber, where it has stronger guardrails. The company said that although Opus 5 is not trained specifically for cybersecurity tasks, the model naturally improved in this area “as a result of becoming more generally capable.”

After OpenAI’s latest models went rogue and autonomously hacked into the servers of AI company Hugging Face earlier this month, Clem Delangue, Hugging Face’s CEO, spoke with Fortune about the delicate balance AI model vendors ought to strike between empowering cyber defenders and not providing too much capability to attackers. Hugging Face had to use a Chinese-built open-source model from Z.ai after an unnamed frontier model from a U.S. company refused Hugging Face’s requests.

“Closed model APIs have guardrails that flag and refuse a lot of legitimate security work, because analyzing an attack looks a lot like preparing one,” Delangue said. “When you’re in the middle of an active incident, you can’t have your tools refusing to examine malicious payloads or getting your account flagged.”

Anthropic said Opus 5 is “the most aligned Opus model… and is the least susceptible to being tricked into misuse.”

The AI vendor is also gearing the new model towards the scientific community, calling Opus 5 its “most capable generally available model for scientific research.” It’s particularly strong in biology tasks compared to its predecessor.

The release announcement also includes testimonials from a law firm, and, of course, software engineering teams, who remain a critical customer base for the company.