Small AI models: why they now fit small businesses best

For a couple of years the default advice was simple: pick the most capable model you can afford and stop thinking about it. That advice has expired. Small models — the ones with "mini", "small" or "lite" in the name — now handle a large share of what a small business actually needs, at a fraction of the price. The odd part is that the reason they should matter to you isn't the one you'll hear from whoever is selling.

The price ladder has stretched wide open

Providers publish rates per million tokens (a token is roughly a chunk of a word; about 750 words per 1,000 tokens). Take Mistral's pricing table — European, which helps if you care where your customers' data ends up. Today the ladder looks like this:

• Mistral Medium 3.5: $1.5 in, $7.5 out per million tokens.
• Mistral Small 4: $0.15 in, $0.6 out.
• Ministral 3 3B, the smallest of the family: $0.1 and $0.1.

So for the text the model writes — which is where the money goes — the mid-tier model costs twelve times the small one, and seventy-five times the tiny one. Google's ladder looks similar: its cheapest general-purpose model, Gemini 3.5 Flash-Lite, sits at $0.30 in and $2.50 out, well below its flagships.

There's also a discount almost nobody claims: the fixed part of your prompt — the instructions, the catalogue, the house rules — can be cached, and Mistral's table bills cached input at 10% of the normal rate. If your agent repeats the same two pages of context on every message, which is the norm, that's another free cut.

Now run the numbers, because the savings argument falls apart

Take a realistic case: an AI receptionist covering WhatsApp and the website, 200 messages a day, roughly 6,000 a month. Each message pulls in around 1,500 input tokens (instructions, catalogue, opening hours, the earlier thread) and produces about 200 out. That's 9 million input tokens and 1.2 million output tokens a month.

• Mid-tier model: $13.5 + $9 = about $22 a month.
• Small model: $1.35 + $0.72 = about $2 a month.

Twenty dollars apart. Less than a decent lunch. If you've been reading this as a cost-cutting exercise, stop: at small-business volumes, the model bill is almost never the problem. What's expensive is a person's time, the integration work, and the messages nobody answers on a Saturday.

So be sceptical when someone pitches a model swap as a savings plan. You'd be optimising the smallest line on your P&L.

What a small model actually buys you

So why do they matter? Three things that never show up on the invoice.

Speed. A small model comes back faster. I won't quote you seconds — they depend on the provider, the length of the text and even the time of day — but the gap is obvious in a live chat, where someone is sitting there watching the screen. In a WhatsApp conversation, answering quickly is part of answering well.

Being able to spend calls. A good agent doesn't make one model call per message; it makes five or six. Classify what the message is about, pull out the details, check the calendar, draft, re-read its own draft, decide whether to escalate. At flagship prices, chaining six steps stings, and you end up cutting exactly the checking steps. At small-model prices you can afford to let the agent proofread itself — which improves results far more than moving up a tier.

Brute-force jobs. Reviewing 4,000 customer records, sorting twelve months of email, pulling data out of 900 scanned delivery notes. With the big model you hesitate; with the small one you kick it off on Friday afternoon and read it on Monday.

The split that works: small model on the door, big model on call

The decision isn't "big or small". It's which part of the job each one does. The pattern that holds up in real businesses looks like this:

message arrives
  → small: what is this about? (booking | price | problem | spam)
  → small: extract details (name, date, reference, amount)
  → routine, and covered by the playbook?
       yes → small drafts and sends
       no  → big drafts it → lands with you for approval

Eight out of ten messages a small business receives are routine: hours, prices, availability, "has my order shipped?". The small model handles that stretch without breaking a sweat. The big one steps in only where it earns its keep: the angry customer, the unusual request, the quote with conditions attached.

There's a side benefit nobody advertises: once the work is split into steps, you can see which step failed. If the agent says something daft, you can tell whether it misread the message or just wrote a bad reply. With one giant call doing everything at once, you're left staring at the screen with nothing to fix.

Where the seams show

Not everything survives the downgrade. Before you switch, test these by hand — this is where small models come apart:

• Long, nuanced conversations, especially an annoyed customer stacking three complaints into one message.
• Many rules at once. If your playbook has ten chained conditions ("if they're on a maintenance contract and the fault is under warranty, unless…"), the small model starts dropping one.
• Long documents: a contract, a tender, a twenty-page report.
• Contradictory information, where the job is deciding which source to trust.

And a warning about labels: the names mislead. This year's "mini" can beat last year's flagship. Don't trust the tier name or the benchmark chart — trust your own messages.

When to change nothing at all

If your agent works and the model bill is under €50 a month, swapping models is rearranging furniture to save the price of a set lunch, at the risk of breaking something that already runs. Don't.

If you're just starting, build version one on the big model. It's far more forgiving of badly written instructions and lets you find out whether the idea works before you start tuning. Moving down later is easy; debugging two problems at once — is it the prompt or the model? — is miserable.

And if the conversation is the sale — high ticket, customer deciding partly on how you treated them — don't economise there. Ten cents of model never covers a lost lead.

How to decide it in one afternoon

1. Pull 30 real messages from the last month. Include the ugly ones: the typo-ridden, the ones that arrive in three fragments, the person shouting in caps at eleven at night.
2. Run them through both models with exactly the same prompt.
3. Label every reply one of three ways: send as-is, needs a tweak, wrong.
4. If the small model produces roughly the same number of "send as-is" replies, switch it for that task and keep the big one for whatever it chokes on.
5. Put a note in the calendar to repeat the test in three months. Prices and models move every few weeks, and what falls short today may clear the bar in the next release.

That's exactly how we build our receptionist: the routine gets resolved by the small model in seconds, anything delicate moves up a tier and lands on your screen as a draft. You approve, it executes.