In short: On June 30, 2026, Anthropic released Claude Sonnet 5. Under a promo running until August 31, agentic tasks on it are 2.5x cheaper than on Opus 4.8 - 2 dollars per million input tokens and 10 per million output. The model plans a long task on its own, works in the browser and terminal, notices its own error and self-corrects. For business the takeaway is twofold: routine that an AI agent finishes end to end got more affordable - but a cheaper AI isn't a ready employee. Without your company's knowledge and rules, it's a new hire with no instructions.
Hello. Today let's translate a piece of news from the language of tech headlines into the language of business. On June 30, 2026, Anthropic released a new model - Claude Sonnet 5. The point isn't "another model shipped," but that it makes an expensive thing noticeably cheaper. We're talking about AI agents - and I'll explain, in plain words, what that is, what it costs, and where the trap for an owner is.
What's new: an agent that finishes the task
First, let's separate two things people often confuse. A regular chatbot is an advisor: you ask, it answers, and you do the rest yourself. An AI agent is a doer: it doesn't just advise, it takes the steps to carry the task to a result.
Picture the difference between a consultant and a repairman. The consultant says: "here's how to fix that tap." The repairman comes and fixes it. Claude Sonnet 5 is positioned as the repairman: per Anthropic and TechCrunch, the model plans a long task on its own, can work in the browser and terminal, and - most importantly - notices its own mistake and self-corrects on the fly, without waiting for you to point it out (Anthropic, 06/30/2026; TechCrunch, 06/30/2026).
Here's a live example from that same release. At Zapier, an agent on the new model ran the whole process itself: it updated customer tiers in Salesforce and messaged the right contacts. That's the chain that used to stall halfway - the agent would slip up somewhere and freeze. Here it carried the job to completion.
What this means for your business: an AI appears that can not only suggest, but run a routine process from the first step to the last - update data, message a client, assemble a report - without you standing over it, prompting every action.
The key number: 2.5x cheaper, not three times
Now about money, because that's the real news. Under the promo running until August 31, 2026, Claude Sonnet 5 costs 2 dollars per million input tokens and 10 dollars per million output (Anthropic, 06/30/2026).
Don't let the word "token" scare you. A token is a small piece of text the model reads or writes. Input tokens are what you feed it (the task, the data, the documents). Output tokens are what it writes back. Think of a taxi meter: it counts kilometers, not trips. Same here - you pay not for "one answer," but for the volume of text read and written.
The key number is simple: the same agentic work on Sonnet 5 comes out 2.5x cheaper than on Opus 4.8. Not three times - exactly 2.5x, and that precision matters, because decisions get built on the difference in figures. Cheaper means processes that used to be uneconomical to hand to AI - too costly to run an agent back and forth - are now within reach even for a small business.
What this means for your business: the barrier to entry dropped. Tasks that once were "let's wait until AI gets cheaper" have moved into "we can run the numbers now." But - and this is the next block - cheaper doesn't mean ready.
The trap: a cheaper AI is not a ready employee
Now the most important part, without which this whole conversation would be advertising, not an honest breakdown. The fact that a powerful AI agent got cheaper doesn't mean you can drop it onto a process tomorrow. A cheaper AI is not a ready employee.
Here's the exact metaphor. Imagine you hired a very capable person. Smart, fast, experienced. But you didn't show them where things are, gave no instructions, didn't introduce them to clients, and didn't explain the company's rules. What will they do on day one? Make mistakes - not because they're dim, but because they don't know your context. An AI agent without your business's knowledge and rules is exactly that: a new hire with no instructions.
The model can plan and self-correct - that's true. But it doesn't know your prices, your services, how you talk to a client, your limits - what you may promise and what you may not, when to stop and call a human. Without that, it answers correctly in form, but not about your business - "out of thin air." And there's a separate question of which of your business data you can safely feed such an agent; I break that down in is it safe to upload customer data to ChatGPT.
What this means for your business: a discount on the model is a discount on the engine, not on a finished car. The engine got cheaper, and that's great. But to make it move your business, you still need a steering wheel, rules of the road and a route - that is, the knowledge and instructions of your specific company.
What turns a cheap AI into a reliable helper
So what's the difference between "bought access to a cheaper model" and "have an agent that actually works." The difference is three things, and none of them is about the price of a token. It's essentially the same story as with a pile of separate tools: buying one more isn't the fix until they're stitched into a system built on your business's knowledge (I broke that down in the piece on the AI zoo: 82% bought AI, only a few win).
First - your business knowledge. The agent has to know your services, prices, typical client questions, your tone. That's its instruction sheet. Without it, even the smartest model answers generically.
Second - rules. What the agent may do on its own, and where it must stop and hand the matter to a person. Those are its boundaries of responsibility - like a new hire who's told clearly: "decide this yourself, but here call the manager."
Third - a check at the end. The agent prepares and drives the process, but the human keeps quality control and the key decisions. A model that notices its own errors is good, but your business shouldn't rest on it grading its own homework.
What this means for your business: these three things are exactly what turn a cheaper AI from an interesting headline into a working tool. At AIevgen CREATOR we build agents this way - on a specific business's knowledge and rules, with a check at the end - so the agent answers about your business, not "out of thin air," and takes off the routine without you losing the steering wheel.
A practical route: where to start
I'd advise you not to try handing everything to the agent at once. The most common mistake is trying to automate the whole business over a weekend. Start with one narrow process and finish it.
First - pick the one process that eats the most time. First replies to clients, data updates, preparing a routine report. One, not five.
Second - write down the knowledge and rules for it: 20-30 typical situations and your answers. What the agent does on its own, and where it asks a clarifying question or calls you.
Third - connect the agent to a single entry point and run a dozen real cases through it. See where it slips, fix the rules - and only then scale to the next process.
What this means for your business: this way you quickly see a real first result on one process, already at the new, lower price, without chaos and without hiring another person - and then you grow the system calmly, one piece at a time.
Summary
So here's the main thing. Claude Sonnet 5 made powerful AI agents noticeably more affordable: agentic tasks became 2.5x cheaper than on Opus 4.8, and the model itself can plan, work in the browser and terminal, and fix its own mistakes. That genuinely lowers the barrier to entry for business. But a cheaper engine is not a finished car. An AI agent starts making money not when it gets cheaper, but when it runs on the knowledge and rules of your specific business, under your control.
Those are exactly the agents we build as part of the AI ecosystem for business service - on your business's knowledge, with clear rules and a check at the end. If you want to figure out which routine to hand the agent first, write the word АГЕНТ (AGENT) to me in Direct and I'll suggest where to start in your case.
And in the comments, tell me: which routine would you hand an AI agent first?
I wish you a peaceful day and a thriving business.
Sources
- Anthropic - Claude Sonnet 5, 06/30/2026 (model release; promo until 08/31/2026 - $2 per 1M input / $10 per 1M output tokens; 2.5x cheaper than Opus 4.8; plans long tasks, works in browser and terminal, self-corrects; Zapier case - updating customer tiers in Salesforce + messaging contacts).
- TechCrunch - Anthropic launches Claude Sonnet 5 as a cheaper way to run agents, 06/30/2026 (positioning as a cheaper way to run AI agents).
FAQ
What is Claude Sonnet 5 and how is it different?
Claude Sonnet 5 is a new model from Anthropic, released on June 30, 2026. It's positioned as a cheaper way to run AI agents: the model plans a long task on its own, works in the browser and terminal, notices its own error and self-corrects. Under a promo running until August 31, 2026, agentic tasks on it are 2.5x cheaper than on Opus 4.8.
How much does Claude Sonnet 5 cost?
Under the promo until August 31, 2026: 2 dollars per 1M input tokens and 10 dollars per 1M output tokens. That is 2.5x cheaper than Opus 4.8, not three times. A token is a small piece of text the model reads or writes; for a business, what matters is that the same agentic work now costs noticeably less.
Can an AI agent complete a task from start to finish on its own?
In the Zapier example, an agent on the new model updated customer tiers in Salesforce and messaged the right contacts - a chain that previously stalled. So technically the agent can carry a process to completion. But it works reliably only when your specific business's knowledge and rules are written into it, with a check at the end.
Does a cheaper AI mean I can let a person go?
No. A cheaper AI agent removes routine and finishes typical processes, but it leaves decisions, quality checks and hard cases to a human. Without your company's knowledge and rules, it's a new hire with no instructions: smart, but it will make mistakes. It's a helper, not a replacement for a person.
How do I start using a cheaper AI agent in my business?
Pick the one process that eats the most time, write down the business knowledge and rules for it (typical questions, what it may do, when to hand off to a person), connect the agent to a single entry point, and test on real cases. One process under control first, then scale. We build exactly these agents on a business's own knowledge.