Anthropic Made the Cheap Workhorse Better. The Ten-Minute Check Most Owners Never Do

Most AI coverage is about the top of the range: the biggest, most expensive model, the one that wins the benchmark headlines. Makes sense. That's the exciting end. It's also not the end that runs your business. The model doing the work behind a small firm's automations is usually the tier below the flagship. That tier just got a real upgrade this week, and it costs you nothing to claim.

On 30 June, Anthropic released Claude Sonnet 5 (the Anthropic release page has the specifics). Two things matter for a business owner. It's better at the work that makes automations useful: planning multi-step jobs, using tools like a browser or a spreadsheet, and running a job through to the end on its own. And it's cheap. $2 per million input tokens through 31 August, $3 after that, against $5 for the top-tier model. On one agentic benchmark it edges the flagship. You're paying less for the capability that's relevant to you.

Why the middle tier is the one that matters to you

Model ranges work a bit like a car line-up. The halo model at the top gets all the press and almost nobody buys it. The cheap runabout does the school run. The sensible mid-range does most of the real driving. In AI, that mid-range is where nearly every practical small-business automation should live. Here's why.

The jobs you'd sensibly automate look like this: sorting enquiries, drafting standard replies, pulling a weekly digest together. Demanding enough that the cheapest model fumbles them. Nowhere near hard enough to need the flagship. The mid-tier is built for exactly that band. Capable enough to be reliable. Cheap enough to run constantly. When it gets better at agentic work, as it just did, the ceiling on “what a small business can automate affordably” moves up. Jobs that needed the expensive model six months ago now run on the cheap one. That's the story, and it never makes a headline.

An example from my own setup. I've got a handful of routines that wake up on a schedule, read what they're pointed at, and leave me a short digest by breakfast. Every one runs on the mid-tier. If I ran them on the top model they'd cost several times as much and do the job no better. The job doesn't need frontier reasoning. It needs to reliably read and summarise what's in front of it. When the mid-tier improves, those routines get sharper overnight and my bill doesn't move. That's the pattern I want for every client.

The problem: most owners don't know what they're running on

Here's where it gets awkward. Ask a business owner which model their automation or their team's AI assistant is using, and the honest answer is usually a blank look. They set it up months ago, picked whichever was the default, and haven't looked at it since. This matters two ways.

If you're pinned to an old model, you're missing free upgrades. A workflow set up a year ago and never revisited is running on last year's capability, quite possibly at last year's higher price, while a newer, cheaper, better model sits one dropdown away. If you're reaching for the flagship for routine jobs, you're overpaying for reasoning you don't use, sometimes by three to five times. Either way, nobody's looking, so nobody catches it. It's the AI equivalent of never checking whether you're on a decent energy tariff. Wasteful, and easily fixed.

What to do once you've looked

If the check turns up an old model on an important job, test the current mid-tier on the same task before you switch. Same input, compare the outputs side by side. Nine times out of ten it's better and cheaper, and the swap is a dropdown, not a project. If it turns up the flagship doing routine work, ask whether the job needs the extra reasoning. Usually it doesn't. Stepping down a tier cuts the cost with no visible drop in quality.

The one case where you keep the flagship is the small number of hard, judgement-heavy tasks where a better answer is worth real money. A complex analysis. A tricky piece of drafting where mediocre isn't acceptable. Pay up for those. Run everything else on the workhorse. The skill is matching the model to the job, and being honest that most of your jobs are ordinary.

The trend worth trusting

Step back and the direction is clear and friendly to a small business. The capable-but-cheap tier keeps getting better faster than it gets more expensive. This week's release is one more step in a pattern that's held all year. The price of “good enough to run your operations” keeps falling. You don't have to chase every release. You do have to look once in a while at what you're running, because the upgrades are free, but only if you claim them.

So this week's task. Find out which model your business is running on. Ten minutes of work. The least glamorous AI job you'll do all month, and probably the best-value one.

Frequently asked questions