Claude Sonnet 5 explained: what Anthropic's cheaper AI agent means for you

If you've heard people rave about "AI agents" and quietly wondered what the fuss is — or worse, seen the bill — this launch is worth five minutes of your time. On 30 June 2026 Anthropic released Claude Sonnet 5, which it calls its most agentic Sonnet model yet. The short version: it does the clever, autonomous stuff that until recently only the big, expensive models could do, but it costs a fraction as much to run. Here's what that actually means for you, in plain English.

A regular chatbot answers your question and stops. An AI agent keeps going: you give it a goal — "research these five suppliers and draft a comparison" — and it plans the steps, opens a browser, runs a terminal, checks its own work, and comes back with the finished job. Think of it as the difference between asking a colleague a question and handing them a task to complete while you do something else.

The catch is that all that back-and-forth costs money. Every step an agent takes sends text to the model and gets text back, and you pay per chunk of that text (a "token"). A ten-step task quietly costs ten times a one-shot answer. Through 2026 a lot of businesses got a nasty surprise when their agent bills landed — the capability was real, but so was the cost. That's the backdrop Sonnet 5 is walking into, and it's aimed squarely at it.

The headline is that Sonnet 5 gets you most of the way to Anthropic's flagship, Opus 4.8, for a lot less. On an agentic coding benchmark — a test of exactly the tool-using, multi-step work agents do — Sonnet 5 scored 63.2%, against 69.2% for Opus 4.8 and 58.1% for the older Sonnet 4.6. On a separate knowledge-work benchmark, Anthropic says Sonnet 5 slightly beat Opus 4.8. In other words, you're not settling for a toy; you're getting something close to the top of the range for the price of the middle.

Here's how the money lines up. Prices are per million tokens (roughly 750,000 words), split into what you send in and what the model sends back out:

So running the same agent on Sonnet 5 instead of Opus 4.8 can more than halve your cost, and during the introductory window it's cheaper still. Sonnet 5 is available on every plan and is now the default model for Free and Pro users, with Max, Team and Enterprise on board too — so most people are already using it without changing a thing.

Being honest with you: there are two things to keep an eye on. The first is a quiet one buried in the fine print — a new tokenizer. A tokenizer is just the way the model chops your text into billable pieces. Sonnet 5 chops differently, so the exact same document can count as anywhere from 1.0 to 1.35 times as many tokens as before, depending on what's in it. That matters because your bill is per token, so a lower headline price doesn't automatically mean a lower invoice.

So the practical advice is simple: if you're running real volume, don't take the sticker price on faith — run your own workload for a day and compare the actual token counts. The second catch is one of context rather than cost: nearly every performance number here comes from Anthropic's own launch post, and the early partner praise is from privileged access, not independent testing. The numbers are plausible and specific, but treat them as the maker's own homework until outside benchmarks confirm them. Anthropic even edited the post on launch day to fix a chart that had undersold Sonnet 5 — a reminder that these figures are freshly minted.

This is the part I find genuinely reassuring. Anthropic made a deliberate choice not to train Sonnet 5 on cybersecurity offence — the skill of writing software that breaks into systems. As a result it's substantially weaker than Opus 4.8 at building exploits, and overall it behaves badly less often than Sonnet 4.6, especially when running as an autonomous agent. That's not the model failing at something; that's the model being built not to do something.

Why does that help you? Because an agent that can browse, click and run commands on its own is exactly the kind of tool you don't want to be a talented burglar. Choosing to leave the lock-picking skills out is a bit like selling a very capable delivery van that simply isn't wired to hotwire other cars — you lose nothing useful, and you sleep better. It's a pointed move, too: it landed the same month the US government pulled two of Anthropic's own models, Fable 5 and Mythos 5, over their cyber capabilities, so Anthropic is quietly making the safer choice a selling point. For an everyday user or a small team, a powerful agent that's boring to a criminal is precisely what you want.