Claude Sonnet 5 Is Out: What It Means for Local Business Automation

Muhammad Shahid·AI-Powered Digital Marketing Consultant

Google Ads Certified

Meta Blueprint Certified

Google Analytics Certified

Updated July 2026

Anthropic released Claude Sonnet 5 on 30 June 2026, and the short version for a business owner is this: the kind of artificial intelligence that powers your automations just got more capable at doing multi-step work on its own, and for a limited window it is cheaper to run. Anthropic is charging an introductory rate of $2 per million input tokens and $10 per million output tokens until 31 August 2026, after which it rises to $3 and $15. If you have been on the fence about an AI tool or agent, that window is a genuine reason to start now. Here is the plain-English read, and what I would actually do about it.

What is Claude Sonnet 5?

Claude Sonnet 5 is the newest mid-tier model in Anthropic's Claude family of large language model systems. Anthropic ships three tiers: Haiku, the small fast one; Sonnet, the everyday workhorse; and Opus, the top-end model for the hardest jobs. Sonnet is the one most people and most tools actually run, because it balances capability against cost. Sonnet 5 is the new version of that workhorse, and Anthropic is blunt about the headline change:

The benchmark numbers back that framing up. On Anthropic's reported results, Sonnet 5 scores well on the kind of tests that measure reasoning and, more importantly for a business, the ability to operate real software. Benchmarks are not the point on their own, but the direction of travel is what matters here.

The OSWorld score is the one I pay attention to. It measures whether a model can drive real software, click through screens, fill forms, move between apps, which is exactly what separates a chatbot that talks from an agent that does. Anthropic reports Sonnet 5 handling that at 78.5 percent, per its launch announcement. That capability is what turns AI from a writing aid into something that can run a workflow.

Sonnet, Opus, or Haiku: which Claude tier should a business use?

For almost every business job, Sonnet 5 is the right default. Anthropic sells three tiers of Claude, and the temptation is to reach for the biggest one. Resist it. The top model costs several times more and only pulls ahead on the hardest reasoning and coding, which is not where a local business spends its AI time. Here is the plain difference between the three.

• Haiku. The small, fast, cheapest model. Good for high-volume, simple tasks like sorting enquiries or tagging messages, where speed and cost matter more than deep thinking.
• Sonnet 5. The everyday workhorse. Strong enough for content, research, and running an automation, at a price you can leave running. This is what most tools use, and what most owners should.
• Opus. The top-end model for the hardest jobs, complex analysis, heavy coding, long chains of reasoning. Worth the premium only when Sonnet visibly struggles.

In practice, here is how I decide which tier to point at a job:

1. Start with Sonnet 5 for anything real. It handles the vast majority of marketing and admin work.
2. Move up to Opus only when you can see Sonnet failing at a specific hard task, not on a hunch.
3. Drop to Haiku for the simple, high-volume steps inside a workflow, to keep the cost down.

What does the pricing actually mean for a business?

For most owners the pricing matters only if you, or a tool you use, pay Anthropic by the token through its application programming interface. If you use Claude through a flat monthly plan, the token price is mostly Anthropic's problem, not yours. But automations and custom agents usually do bill at token rates, so the window is real money for anyone running one. Here is the full picture, straight from Anthropic's pricing documentation.

A worked example, in money you recognise

Say you run a modest automation that reads and writes about 5 million input tokens and 1 million output tokens a month, roughly a follow-up-and-reporting agent for a small service business. At the introductory rate that is about $20 a month. At the standard rate it is about $30. So the window saves that workload around $10 a month, and, more to the point, it shows how cheap capable AI now is: a machine that follows up on every lead and drafts your weekly report for the price of a couple of coffees. In Australian dollars that is roughly 30 versus 45 a month; in dirhams, about 75 versus 110. The token cost was never the thing stopping you.

Why “most agentic” matters to a local business

“Agentic” is the word doing the real work in this launch. It means the model is better at running a job end to end, what computer scientists call an intelligent agent, rather than answering one prompt and stopping. A more agentic model plans the steps, uses tools, checks its own output, and keeps going. That is the difference between AI that drafts a reply and AI that follows up on the lead, books the call, and logs it in your system.

For a local business, that reliability is the whole game, because an agent you cannot trust to finish a job is just extra supervision. The jobs that get more dependable as models get more agentic are the boring, valuable ones:

• Speed to lead. Following up on every missed call and web enquiry within minutes, across text and email, without dropping one on a busy day.
• Reviews. Asking happy customers at the right moment and drafting a reply to each review that lands.
• Reporting. Pulling calls, leads, spend, and rankings into one plain weekly summary.
• Content drafts. Turning one job into a week of posts in your voice, for you to approve.

None of that is new in theory. What changes with each model is how often the agent finishes the job cleanly without a human catching an error. If you want the fuller picture of how these systems run, I broke it down in what AI agents for marketing actually are and, for the mechanics, how the agent loop works.

Should you switch, or wait?

Here is the honest advice, and it is not the exciting one: do not switch tools to chase a model. You almost never buy a model directly. You use an app or an agent built on top of one, and a good provider swaps the model underneath for you when a better one ships. Most people using the Claude app are already on Sonnet 5 without lifting a finger, because Anthropic made it the default. The same pattern holds across the field, whether the engine underneath is from Anthropic, OpenAI behind ChatGPT, or Google behind Gemini.

What a launch like this should change is not your tool, but your timing. A faster, cheaper, more capable model raises the floor for every product built on it, and the introductory pricing lowers the cost of trying one right now. So the smart move is not to migrate. It is to finally run the automation you have been meaning to test, while the underlying generative artificial intelligence is at its cheapest. If you are weighing Claude against the alternatives for actual work, I compared them in ChatGPT Work vs Claude Cowork.

What I would do before 31 August

If I were an owner reading this, I would treat the window as a deadline to run one small test, not to overhaul anything. Here is the order I would work in.

A model launch is a headline. Whether it helps your business comes down to one unglamorous question: did you point it at a job that was costing you money, and did you check its work? That is the same discipline behind any good marketing automation, model or no model. If you want a second pair of eyes on which job to automate first, that is exactly what I look at in a free audit.

Frequently asked questions

These are the questions owners ask me most about Claude Sonnet 5. The same answers are embedded in this page's FAQ schema for search engines and AI answer engines to read.