Yuval Avidani
Author
"The most agentic yet." That's the description Anthropic gave today, June 30, 2026, to their new model, Claude Sonnet 5. And I know that word, "agentic," sounds like just another buzzword thrown onto a slide. But it turns out this time it actually holds up, and I want to break down why.
Let's start with the funny name: the model's internal codename is Fennec, as in the fennec fox, that tiny desert fox with the giant ears. A creature that's basically all sensors, moving fast through the desert. Coincidence or not, it's a pretty accurate description of what this model is supposed to be.
So what does "agentic" even mean, without the bullshit
Let's break it down. An "agentic" model doesn't just answer us, it plans steps, runs tools on its own, and checks whether it succeeded. Think of the difference between an employee you need to tell every single action ("now open the browser, now search, now copy this"), versus an employee you tell "find me the three cheapest flights to Prague and make me a table," and they go do the whole thing themselves.
The tools Anthropic mentions are browser and terminal. Browser is obvious: browsing websites, reading, clicking. Terminal is that black window where you run commands on a computer, the thing developers live inside all day. When a model can operate both of these on its own, it's no longer "chat", it's something much closer to an employee.
The definition worth remembering: Claude Sonnet 5 is Anthropic's mid-priced model, one that plans and runs tools on its own (agentic), built for people building automations and agents without burning through their budget. And the exciting part is that last bit: without burning through your budget.
Why "cheap" is the story here, not "smart"
To understand why the price is what's exciting me, you need to understand how you actually pay for models. When we send text to a model, it doesn't see words, it breaks them down into tokens. A token is a small chunk of text, roughly a word or half a word. Think of it like language Lego bricks: "hello" might be one token, and a long word might break into several.
Below is a real engine that shows this: let's write some text and see how it breaks apart:
Now, per-token pricing is simple: you pay based on the number of tokens going in (what we sent) and coming out (what the model replied). Think of it like a taxi meter: the longer the ride, the more you pay.
And here are Sonnet 5's numbers. Introductory pricing, through August 31, 2026: $2 per million input tokens, and $10 per million output tokens. After that date it jumps to $3 and $15. Its API identifier is claude-sonnet-5, it's the default model in the Free and Pro plans, and it's also available on Max, Team and Enterprise.
Why does this matter to us? Because when you build an agent, the model doesn't run just once. It runs in a loop: plan, try, fail, fix, try again. Every one of those rounds is more tokens, more money on the meter. With an expensive model, an autonomous agent quickly turns into a black hole in your wallet. With a cheap model that also knows how to work on its own, suddenly the math works.
Fairness check: it's not the most accurate, and that's fine
It's important for me to be fair here, because it's easy to get carried away. Anthropic themselves position Sonnet 5 as the cheap alternative, an alternative to their own Opus, to OpenAI's GPT-5.5, and to Google's Gemini Pro. Notice the word "alternative", not "winner".
And in the fine print, Anthropic say it themselves: Opus 4.8 is still more accurate on certain tasks. Meaning, if we're doing something where every mistake is costly (legal analysis, critical code, a medical decision), we might actually want to pay more for the more accurate model. That's a different approach, not an inferior one. And GPT-5.5 and Gemini Pro also have their own advantages in different areas. There's no single "winner" here, there are different tools for different jobs.
How does the model even "decide" what the next step is
A lot of people ask me how a model can "plan" if all it's doing is guessing words. So here's the intuition: a language model, at its core, predicts the most likely next word (token). Over and over. It sounds too simple to work, but when you do that billions of times over a massive amount of text, you get behavior that looks like planning.
There's a real engine below that shows exactly this: let's start a sentence and it'll guess the next word:
"Agentic-ness" is built on top of that: instead of the model predicting only words, it also predicts "actions": when to open a browser, when to run a command, when to check the result. In my view, that's the interesting move here: it's not that the model got smarter, it's that they taught it to work in steps instead of in one big leap. Exactly like the difference between someone who just blurts out an answer versus someone who stops, thinks, and double-checks themselves.
What this means if you're building, and why an IPO is in the air
Now to my angle on this. If we're building automations, agents, or any workflow where a model runs on its own, Sonnet 5 is exactly the kind of tool that changes the math. Not because it's the smartest model in the world, but because the combo of "cheap + independent" is exactly what lets you run things that until now were too expensive to be worth it.
Think about it: a few months ago, this level of autonomy required the big, expensive models. Now the same work runs on a mid-tier model that costs a fraction of the price. This isn't magic. It's the normal process in this industry, where a capability that used to be premium turns into a cheap default in a short amount of time.
And there's business context here we can't ignore. Anthropic are in a race toward an IPO, an initial public offering, meaning the moment a private company first offers shares to the public on the stock exchange. On top of that, they just launched a new product called Claude Science. Aggressive introductory pricing, a cheap default model, and a new product right before an IPO: these aren't coincidences. Important for me to say: this isn't investment advice or financial guidance, just a factual description of what's happening.
Bottom line, as I see it: Sonnet 5 doesn't beat anyone on raw power, and Anthropic aren't claiming it does. The story is that they dropped the price of "a model that works on its own" to a point where a lot more people can suddenly build real agents. The limitation is clear: for the most critical tasks, more accurate models still exist, and it's on us to figure out where our own threshold is. So I'll ask you: with a cheap, agentic model in hand, what's the first thing you'd let it run and do on its own?