Just a few months ago, having an artificial intelligence that could genuinely browse the web, open a terminal, and fix a bug on its own felt like a luxury reserved for the most massive, expensive models on the market. If you wanted a true “agent”—an AI that doesn’t just talk, but actually acts on your behalf—you had to pay a premium.

That math is rapidly changing. Yesterday, Anthropic rolled out Claude Sonnet 5, and it’s arguably the company’s most significant push yet toward making autonomous, “agentic” AI accessible to everyday developers and businesses.

To understand why this matters, you have to look at Anthropic’s product lineup. Sonnet has traditionally been the company’s middle child. It sits comfortably between the lightning-fast, lightweight Haiku and the heavy-hitting powerhouse Opus. Historically, if you wanted an AI to reliably string together complex multi-step tasks without getting confused or stalling out halfway, you had to bite the bullet and upgrade to an Opus-class model. But Sonnet 5 is actively blurring that boundary. According to Anthropic’s internal benchmarks, this new model is breathing down the neck of the much larger Opus 4.8 in critical areas like reasoning, coding, and tool use.

Let’s demystify what “agentic” actually means in this context, because the tech industry loves its buzzwords. We’re moving past the era of the simple chatbot. An agentic model is essentially given a goal, and it figures out the steps required to get there. It can use a web browser to look up missing information, interact with a terminal to run code, realize it made a mistake, and correct itself—all without a human needing to hold its hand or prompt it at every turn.

Early testers are already painting a picture of what this looks like in practice. Software engineers report handing Sonnet 5 a ticket to investigate a bug, only to watch the AI independently write a test to reproduce the issue, implement the fix, and then verify that the bug was actually gone. Others in the legal and insurance sectors are using it to completely automate multi-step workflows like submitting intake forms and generating loss runs, tasks that normally require a human to manually bounce between different software systems.

While this level of autonomy is impressive, the real story here is the economics. The foundational bottleneck for AI startups right now isn’t just capability; it’s the cost of API calls. Anthropic is launching Sonnet 5 with an introductory price of $2 per million input tokens and $10 per million output tokens through the end of August 2026. Even when the standard pricing kicks in this fall—$3 for inputs and $15 for outputs—it will still severely undercut the cost of using Opus 4.8, which sits at a hefty $5 and $25, respectively.

What makes this particularly interesting is how Anthropic is letting developers turn a dial on something called “effort levels.” If you give Sonnet 5 more time and compute to think through a problem, it can actually match Opus 4.8 on certain complex tasks. It gives businesses a sliding scale to balance cost against performance, rather than forcing them into a rigid, expensive pricing tier just to get a job done.

Of course, whenever an AI gets better at writing code and navigating systems autonomously, the security community starts to sweat. Anthropic’s safety evaluations highlight a fascinating tension inherent in building smarter models. On the whole, Sonnet 5 is safer than its predecessor. It hallucinates less, is less prone to sycophancy (agreeing with the user just to be polite), and is better at refusing outright malicious requests.

However, its generalized boost in intelligence means it’s slightly more capable of figuring out software vulnerabilities. In tests developed alongside Mozilla, Anthropic found that while Sonnet 5 still can’t build a working cyber exploit from scratch, it showed a slightly higher rate of “partial success” than older models. Because of this slight bump in capability, Anthropic is playing it safe and launching the model with real-time cyber safeguards enabled by default to detect and block dangerous usage.

We’re clearly entering a new, highly pragmatic phase of the generative AI boom. The novelty of a machine that can write a passable sonnet or generate a recipe has worn off; the focus now is entirely on utility and execution. Companies don’t just want an AI that can draft an email. They want an AI that can update a CRM, write the launch announcement, and hit send, all in one seamless workflow. With Sonnet 5, Anthropic isn’t just releasing a smarter chatbot. They are making the foundational layer for real, multi-step automation cheap enough and reliable enough to actually build a business on.