Anthropic released Claude Sonnet 5 on June 30, 2026, and the headline most people will repeat is simple: new model, cheaper. The part that actually matters for a business is quieter. Work that needed Anthropic’s most expensive model a few months ago now runs on a mid-tier one. That changes the math on what you can hand off.
This is a plain-language look at what the new model is for, what it is good at, how it stacks up against the latest Opus, and how to decide which one a given job needs. No hype, no jargon, just what it means for the work you actually do.
What Claude Sonnet 5 is
Anthropic calls it their most agentic Sonnet yet. In practice that means it does not just answer questions, it acts. It makes a plan, uses tools like a browser or a terminal, and runs multi-step work on its own without you steering every turn.
A bit of background helps here. The Sonnet line is the middle tier of Anthropic’s models, sitting between the small, fast Haiku and the large, powerful Opus. For a lot of teams the agentic era really started with Sonnet, the 3.5, 3.6, and 3.7 versions were the first that could handle real coding and tool use. More recently the biggest jumps came from the pricier Opus models. Sonnet 5 is Anthropic pulling that capability back down into the affordable tier.
It is now the default model on the free and pro plans, and it is available to Max, Team, and Enterprise users, inside Claude Code, in Cowork, and through the API as claude-sonnet-5. In short, wherever you already use Claude, this is what you are using now unless you pick something else.
What agentic actually means for your work
Agentic is a word the industry throws around loosely, so here is the practical version. A normal model waits for you, answers, and waits again. An agentic model takes a goal, breaks it into steps, does the steps using real tools, and keeps going until the job is done or it gets stuck.
The difference shows up in how much babysitting a task needs. Anthropic’s early testers kept describing the same thing: the model finishing work where older versions would stop halfway. One detail came up again and again, Sonnet 5 checks its own output without being told to. For anything you would normally hand to a junior assistant and then re-check yourself, that self-checking is the line between real help and more work.
What it is good for, in real use cases
Here is where it earns its place, grouped by the kind of work most businesses actually have.
Coding and technical work. This is its strongest area. Testers ran it against real pull requests and it carried each one to a tested, verified result on its own. One asked it to investigate a bug, and without being prompted it wrote a test to reproduce the problem, implemented the fix, then confirmed the bug came back without the change. That is the workflow of a careful engineer, not a code generator.
Operations and automation. The kind of multi-step task that used to stall is now finishable. One tester handed it a two-part job, update the customer records in the CRM, then send a launch announcement to enterprise contacts, and it ran the whole thing end to end. For routine internal automation, that is the difference between a tool and a teammate.
Research and analysis. It reasons in tighter steps, which means it gets to an answer with less wandering. Teams working with live data reported noticeably faster time to insight. Legal teams pointed to research and analysis as the clearest wins.
Everyday professional work. Drafting, summarizing, cleaning up data, writing first versions of documents, the daily grind of knowledge work, at a quality that used to cost a lot more per task.
The common thread is delegation. The list of tasks you can hand over and trust to come back finished just got longer.
How it compares to Opus 4.8
Opus 4.8 is still Anthropic’s top model. It is the most accurate, the strongest on the hardest problems, and the one to use for sensitive or high-stakes work, including most cybersecurity tasks. It is priced at $5 per million input tokens and $25 per million output tokens.
Sonnet 5 lands close to Opus 4.8 on a lot of tasks for a fraction of the cost. Standard pricing is $3 per million input tokens and $15 per million output, and there is an introductory rate of $2 and $10 running through August 31, 2026. Anthropic’s own performance charts make the point cleanly: where the previous Sonnet fell well short of Opus, Sonnet 5 and Opus 4.8 now cover one continuous range. Sonnet 5 gives you the lower-cost option, Opus the higher-accuracy one, and there is no longer a big gap between them.
There is also a control worth knowing about. You can set an effort level, telling the model how hard to work on a task. Higher effort costs more and buys more accuracy. So even within Sonnet 5 you can dial the balance up for a tricky job and down for a routine one.
When to use Sonnet 5 versus Opus 4.8
The decision is simpler than the model names suggest.
Use Sonnet 5 as your default for almost everything. Day-to-day work, coding, automation, research, content, customer workflows. It is the workhorse.
Reach for Opus 4.8 when a task genuinely needs the ceiling. The trickiest reasoning, the highest-stakes output where a mistake is expensive, and security work where Anthropic specifically recommends Opus. If you are not sure a job needs Opus, it probably does not.
Use the effort dial to fine-tune. Before jumping to the more expensive model, try Sonnet 5 at a higher effort level. Often that closes the gap at a lower price.
What it costs, in plain numbers
For most people on the Claude apps, Sonnet 5 is included, it is the default on the free and pro plans. The pricing matters most if you build on the API or run automation at volume.
The numbers again: $2 per million input tokens and $10 per million output during the introductory window through August 31, then $3 and $15 after. Opus 4.8 runs $5 and $25. So at standard rates Sonnet 5 is roughly 40 percent of the input cost and 60 percent of the output cost of the top model, for performance that now overlaps with it.
One honest footnote. Sonnet 5 uses an updated way of counting text, so the same input can map to slightly more tokens, somewhere between the same and about a third more depending on the content. Anthropic set the introductory pricing so the switch is roughly cost-neutral. The practical takeaway is that running agents and automations got cheaper, not more expensive.
Is it safe to hand real work to
This matters more as the model does more on its own. A model you let touch your tools and data needs to behave.
Anthropic reports that Sonnet 5 is better behaved than the version it replaces. It shows lower rates of hallucination, less sycophancy, the tendency to just agree with you, and a lower overall rate of undesirable behavior. It is better at refusing bad requests and at resisting prompt-injection attempts, where someone hides instructions in content to hijack the model. It ships with cybersecurity safeguards on by default, and it is deliberately weaker at offensive security tasks than Opus.
None of that means you set it loose and walk away. Any model running on something that matters needs a human checkpoint. But as a supervised workhorse, it is a step in the right direction, not a step you have to worry about.
How to actually start
If you want to put this to work rather than just read about it, a simple path:
Pick one repeatable task you already do, ideally a multi-step one that eats time. Run it through Sonnet 5 a few times and compare the output to what you would have produced. Keep a human checkpoint on anything that ships or touches a customer. Then measure the time it saved. The introductory pricing window through August is a good reason to test now rather than later.
The right question is not whether the model is impressive. It is which of your tasks you can now trust it to run.
Quick answers
Is it free? Yes, Sonnet 5 is the default model on the free and pro Claude plans.
What does it cost to build on? Through August 31, 2026, $2 per million input tokens and $10 per million output. After that, $3 and $15.
Should I switch from Opus 4.8? Make Sonnet 5 your default and keep Opus for the few tasks that need top accuracy or are security-sensitive.
Is it safe for automation? It is safer than the previous Sonnet and ships with safeguards on, but keep a human checkpoint on anything important.
Where can I use it? The Claude apps, Cowork, Claude Code, and the API, plus the major cloud platforms.
The bottom line for your business
If you use AI-assisted work at all, Sonnet 5 is the new default: most of the capability of the top model at roughly half to a third of the price. Make it the workhorse, keep Opus 4.8 for the handful of tasks that need the ceiling, and use the introductory window to test it on one real workflow you care about.
If you want help putting models like this to work in your marketing or operations, that is the kind of system we build.