Google has been quietly building a new bottom rung of AI models: cheap, fast ones built to grind through enormous piles of simple work. Gemini 3.5 Flash-Lite runs at about 30p per million words in and £2 per million words out; the newer 3.6 Flash went to a stable release this month. Google is explicit about what these are for: a fleet of ten of them costs a fraction of one flagship doing the same jobs. For an online retailer with thousands of products, or any business drowning in repetitive text work, this changes the maths on jobs that were never worth a person's time. But running everything through the cheap model is a mistake. The pattern that works is two models: a cheap one that does the grinding, and a good one that checks the tricky bit.
Most of the AI news you read is about the top of the range: the biggest, dearest, cleverest model, and which lab is briefly ahead. That's the exciting part, and it's mostly irrelevant to how a real business saves money. The useful shift this year has happened at the bottom of the range, and it's had almost no coverage because "cheap model gets slightly cheaper" doesn't make a headline. It should, if you run anything at volume.
Here's the concrete version. Google now sells a tier of model built for what it openly calls "subagent" work: a Gemini model whose whole job is to be the worker, not the brain. Gemini 3.5 Flash-Lite is priced at roughly $0.30 per million tokens of input and $2.50 per million out, and this month its sibling, Gemini 3.6 Flash, reached a stable, production-ready release with better token efficiency again. Google's own pitch is unusually blunt: this is the model priced so that a fleet of ten costs a fraction of a single flagship doing the same tasks. It is built for "high-volume, low-reasoning work": search, document processing, translation. In plain English: reading, sorting and rewriting, at industrial scale, for pennies.
Why "cheap and dim" is exactly what a lot of jobs need
There's an instinct, when you're paying for AI, to want the best model on everything. It feels responsible. It's usually wasteful. Most of the repetitive text work in a business does not require a genius; it requires a diligent worker who never gets bored. Rewriting a product description into a consistent house style. Tagging an incoming email so it lands in the right queue. Pulling the five fields you care about out of a scanned form. Turning a supplier's messy spreadsheet into your format. None of these is hard. All of them are endless. And paying flagship prices to do them is like hiring a barrister to file your post.
A worker model changes the calculation on precisely this tier of task. When the per-job cost drops far enough, work that was never worth doing, because a person would take a week and a good model would cost more than the value it created, suddenly clears the bar. That's the real unlock. It isn't that the cheap model does anything the expensive one can't. It's that it makes a whole category of dull, high-volume jobs cheap enough to be worth automating at all.
The e-commerce version, because that's where it bites first
Take an online jewellery retailer with several thousand products, a business whose shape I know well. Each product ideally wants a clear, consistent description, the right attributes filled in (metal, stone, carat, size), alt text on the images so search and shoppers' assistants can understand them, and a tone that sounds like the brand rather than a database. Doing that by hand across a few thousand items is a job nobody ever finishes; it's the task that's permanently three-quarters done and quietly costing sales on the quarter that isn't.
A worker model eats this. Point it at the catalogue, give it the house style and the rules, and let it grind through every product, cheaply, overnight. The economics that made this a "one day, maybe" project turn it into a weekend one. And the same pattern applies far beyond retail: a property firm normalising thousands of listing descriptions, an accountancy practice extracting figures from a year of scanned receipts, a support team triaging a flood of inbound messages into the right buckets. Anywhere the work is high-volume and low-judgement, the worker model is the right tool.
The mistake: running everything through the cheap model
Here's where owners go wrong once they've seen the price. They get excited, and they route everything to the worker model, including the 10% of cases that need judgement. Then the cheap model, doing its honest best on a job above its pay grade, produces a description that's subtly wrong, or mis-files the one email that mattered, and the whole thing gets a reputation for being unreliable. The model didn't fail. The design did.
The pattern that works is two models, not one. The cheap worker does the grind, all of it, fast, at scale. A better model sits behind it as a checker, and it only ever looks at the small slice that's flagged as uncertain or high-stakes. On a catalogue, that might mean the worker writes every description, and the good model reviews only the highest-value pieces and anything it wasn't confident about. You pay pennies for the 90% that's routine and flagship rates for the 10% that's worth it. That's the whole architecture, and it's why "subagent" is the right word: the cheap model works for a smarter one, not instead of it.
What I'd do first
I wouldn't start by rebuilding anything. I'd start by finding the one job in your business that is high-volume, repetitive, and low-judgement. The one that's always three-quarters done. Just one. Then I'd run a small, cheap pilot of the worker model on a slice of it, with a human eyeballing the output, to answer a single question: is the quality good enough that I'd trust it on the routine 90%, with a checker on the rest? If yes, you've found something worth scaling. If no, you've spent a few pounds learning that this particular job needs more judgement than you thought, which is also worth knowing.
The temptation with cheap tools is to use them everywhere at once. The discipline that makes them pay is the opposite: pick the dullest, highest-volume job you've got, prove the two-model pattern on it, and only then look for the next one. The worker model isn't exciting. It's the least exciting thing to happen in AI this year. It's also, for a business that runs at volume, quietly the most useful.