Anthropic just released Claude Opus 5, and it comes with a feature designed to solve the problem enterprise customers actually complain about: expensive AI bills. According to the company, Opus 5 includes a control that lets users toggle how much effort the model expends on a given task or prompt, with low, medium, or high settings. The intent is straightforward, at least in theory. Balance between cost and capability, so teams can spend more when they truly need performance and spend less when they do not.

Opus 5 also tries to reduce the “operational tax” of getting good outputs. Anthropic says it is easier to use than its predecessors, requiring less back and forth. It verifies its own work and can recover from errors without intervention, which matters because in practice, iterative prompting and correction are part of why token bills climb. If you have ever watched a workflow turn into a prompt ping-pong match, you know why a feature like this is not cosmetic.

This release lands in an unusually fast-moving stretch for Anthropic. It is the fourth model the company has released in less than two months, after Mythos 5, Fable 5, and Sonnet 5 in June. But that speed comes with baggage, because Fable 5 was controversial enough that the U.S. government temporarily imposed export controls on it after Amazon researchers reported they could bypass its safeguards. Anthropic pulled Fable 5 from the market on June 12, then re-released it on June 30 after fortifying its security.

Enterprise customers and developers then raised another practical issue that hits directly at budgets: burn rate. In the case of Fable 5, critics said it tended to use a high number of tokens to complete tasks. That meant users could blow through token budgets or run up large bills, turning “powerful model” into “expensive workflow.” Anthropic’s positioning for Opus 5 is basically an attempt to split the difference. The company says Opus 5 “comes close” to Fable 5’s capabilities, but at half the price. And it still draws a line for advanced needs. Anthropic recommends Fable 5 for more advanced projects, including those where the model may handle tasks autonomously for days at a time.

The cost conversation is now a race with other frontier labs, not a single-company concern. OpenAI, for example, emphasized economical token usage when marketing its most recent model, GPT-5.6, released on July 9. That matters for anyone running procurement, budgeting, or vendor risk processes. The battleground is not only intelligence. It is unit economics. Tokens are the measurement and the bill. The source notes tokens are units of information an AI model processes, roughly equivalent in English text to about a word and a half, and companies tend to bill for model usage by tokens. A toggle that maps effort to cost is, in that sense, a governance feature for spending, not just a UX feature.

Still, Opus 5’s release is also about safety, and safety is where the industry’s incentives get messy. Anthropic says Opus 5’s safeguards are similar to those of Opus 4.8 in most ways, except in cyber, where it has stronger guardrails. The company frames Opus 5 as “the most aligned Opus model” and says it is the least susceptible to being tricked into misuse. It also says Opus 5 is not trained specifically for cybersecurity tasks, but the model naturally improved in this area “as a result of becoming more generally capable.”

Why the focus on cyber? Because the line between “defensive security” and “helping attackers” is thin, and guardrails can break legitimate work. Fortune also includes context from Clem Delangue, the CEO of Hugging Face, who spoke with Fortune about the delicate balance vendors should strike. He described how closed model APIs with guardrails can flag and refuse legitimate security work because analyzing an attack can look like preparing one. He also highlighted a real operational constraint: when you are in the middle of an active incident, you cannot have your tools refusing to examine malicious payloads or getting your account flagged.

Anthropic adds another layer to this balancing act. As with Fable 5, when Opus 5 declines a user request due to a safety concern, the API automatically falls back to another model, so the user gets an answer rather than a raw error. That design choice is important because it changes what users experience. Instead of an immediate refusal that can halt incident response or development workflows, the system steers you toward a different model path.

Finally, Anthropic is not only targeting enterprise developers and software engineering teams, it is also pitching Opus 5 to the scientific community. The company calls it its “most capable generally available model for scientific research,” and says it is particularly strong in biology tasks compared to its predecessor. The release announcement includes testimonials from a law firm as well, underscoring that legal teams are watching these systems for productivity gains and controllability.

Strategically, Opus 5 is a signal to the rest of the market: vendors are now competing on three things at once. Output quality, spend per task, and how safely the model behaves when users push into sensitive domains. For executives and boards, the stakes are clear. If you buy models based on raw capability alone, you risk token burn and unpredictable workflow failures. If you buy them based only on cost, you risk underperformance that forces more iterations, ironically increasing overall spend. The cost-capability toggle, combined with cyber guardrail adjustments and fallback behavior, is Anthropic’s attempt to make that tradeoff manageable in the real world.