Your Free AI Assistant Is About to Get Smaller

A candle maker writes her product copy in the Gemini app on a Thursday morning. She pastes in a supplier price list to check the maths, turns four rambling customer emails into one clear FAQ entry, and asks for a German version of a listing before lunch. She pays nothing for any of it. On 9 October 2026 that account changes shape, and the change is easy to miss because the thing being taken away is something she never chose in the first place.

This is a small story with a long tail. The mechanics of how an AI vendor re-sorts capability into price tiers is going to repeat, and the useful part is not the outrage. It is learning to tell, quickly and on your own numbers, whether a downgrade touches the work you actually do.

What changes on 9 October, exactly?

Google is restricting which Gemini models each subscription tier can reach. The Gemini free tier is not disappearing, it is being narrowed: an account with no AI plan keeps Flash-Lite, the smallest of the three current models, and loses the ability to switch to Flash or Pro.

The tiers below it shift too. Notebookcheck reported the 9 October date and the per tier breakdown, noting that Google AI Plus at $4.99 a month keeps Flash-Lite and Flash but loses Pro, while AI Pro at $19.99 keeps all three models and picks up Deep Think. The Decoder set out the same tier structure and read it as Google moving closer to a conventional freemium shape, with the heavier models reserved for people who pay.

One detail is worth checking yourself, because it shows how early this still is. On 5 October 2026, the model access table published on Google's own Gemini Apps help page for limits and upgrades still listed Flash-Lite, Flash and Pro as available on every plan, including accounts without a subscription. The dated announcement lives in a separate help article. If you open the table today and conclude nothing is happening, that is why.

Google's description of the models is the clearest guide to what each one is for. The help page calls Flash-Lite an efficient workhorse model designed for speed, suited to everyday tasks such as summarisation and brainstorming. Flash is described as balancing speed and reasoning across problems from simple to complex. Pro is positioned for complex maths and coding, with deeper understanding across text, files, images and video, and the page warns that Pro responses generally take longer.

Read that back as a shopkeeper rather than a developer. Summarising and brainstorming is a fair description of most of what a one person business asks an assistant to do all day.

Which of your jobs actually needs the bigger model?

Most routine shop work sits inside what Flash-Lite is built for. The jobs that degrade are the ones with long inputs, several dependent steps, or arithmetic that has to stay right across many rows.

The table below is my own mapping of common one person shop tasks against the constraint most likely to bite. The input sizes are rough working figures for planning, not measurements from Google.

For a small shop, the split in that table is the whole decision. If your week is the top three rows, the 9 October change is close to a non event for you. If it is the bottom three, you were already fighting the tool and the change makes that fight harder.

Why does the context window matter more than the model name?

Because the thing that stops a task is usually how much you can paste in, not how clever the model is once it reads it. The same Google help page sets the context window at 32k tokens for an account without an AI plan, 128k on AI Plus and 1 million on AI Pro and AI Ultra.

That ladder went almost unmentioned in the coverage of the tier change, which concentrated on model names. It deserves more attention, because it is the number a merchant runs into first and the one that produces the most confusing failure. A model that cannot see your whole document does not tell you it only read part of it. It answers confidently about the part it saw.

Tokens are not words, but for English prose a token is commonly treated as roughly three quarters of a word for planning purposes. On that basis, 32k tokens is somewhere near 24,000 words of input, shared between what you paste, the conversation so far and the answer coming back. A single detailed supplier catalogue, a year of reviews for one product, or a long chat you have been adding to since Tuesday can each eat it.

The gap between tiers is the striking part. Moving from no plan to AI Plus multiplies the window by four. Moving to AI Pro multiplies it by roughly thirty. Whatever you think of the pricing, the capability being sold at the top tier is mostly the ability to hand the thing a big pile of your own material at once.

How do you test whether the downgrade hurts you?

Pick three jobs you genuinely run, save the exact inputs, and score the outputs against a rubric you wrote before you read them. Anything less turns into an argument with yourself about whether the answers feel worse.

The method matters more than the tool, and it transfers to every assistant you will rent after this one.

Choose three real jobs. Not a demo prompt. The three things you did last week that an assistant touched. Write them down as tasks, not as prompts.

Freeze the input. Copy the actual product notes, the actual customer email, the actual spreadsheet extract into a file you keep. A test you cannot repeat identically is not a test.

Write the rubric before you look. Three to five checks per job, each answerable yes or no by someone who is not you. For a product description: does it state the material, does it give a dimension, does it avoid claims you cannot support, would you put your name on it. Binary checks stop you grading on vibes.

Run it twice, on either side of the date. Same input, same rubric, once this week and once the following week. Keep both outputs.

Count only the failures that cost you time. A sentence you would have edited anyway is not a regression. A missing dimension you have to go and look up is.

Three jobs scored this way takes under an hour and produces something you can act on: either the free tier still does your work, or you know precisely which task broke and what fixing it is worth per month.

Flash-Lite is the floor for everybody, not only free accounts

There is a line on the Google help page that reframes the whole tier debate. It says that if you have a Google AI subscription and reach your limit, you can continue your conversation with Flash-Lite.

So the smallest model is not the free tier's model. It is the model every tier lands on when its allowance runs out. Paying buys a larger allowance before you get there, not a permanent seat somewhere else. For a shop that works in bursts, a Sunday evening spent rewriting forty listings, that distinction is the one that decides whether a subscription is worth it.

The same page adds a second mechanism that behaves differently from a plan restriction. It says that for users without a Google AI plan, some features that need more compute, such as Deep Research, may be unavailable during periods of high demand, and that limits for accounts without a plan may be reduced before those with one. That is capacity rationing rather than tiering, and it has a nastier shape: it fails when everyone else is busy, which tends to be exactly when you are busy too.

Usage limits themselves are metered on compute rather than message count. The page says limits factor in the complexity of your prompt, the models and features you use and the length of your chat, and that the allowance refreshes every five hours until you reach a weekly limit. A long conversation you keep returning to is therefore more expensive than several short ones, which is the opposite of how most people assume a chat window works.

What does a smaller model actually get wrong?

Not spelling, and not fluency. A smaller model writes clean sentences; what it loses is the ability to hold several constraints in mind at once and to notice when it has run out of information.

Four failure modes show up repeatedly in shop work, and they are worth learning to spot because none of them announce themselves.

Quiet truncation. You paste more than the window holds. The model answers about the portion it received and says nothing about the rest. This is the one that causes real commercial damage, because the answer looks complete. The defence is to ask a question whose answer you already know and that depends on material near the end of what you pasted. If it cannot answer that, it did not read that far.

Constraint drop. You give six requirements for a product description and get five. Smaller models tend to satisfy the requirements stated earliest and most emphatically, and silently shed the rest. Numbering your constraints and asking the model to list them back before writing catches most of this.

Arithmetic that is nearly right. Margin calculations across a table, unit conversions, totals with a discount applied in the wrong order. The output is formatted like a spreadsheet and carries the authority of one. Any figure you would put in an invoice or a price list needs checking in a tool that actually computes, not a tool that predicts text that looks like computation.

Confident filling of gaps. Asked to draft a policy from incomplete notes, a weaker model is more likely to invent a plausible clause than to flag the gap. We have written at length about how an assistant invents product details that were never in your catalogue, and the same mechanism applies to policies, specifications and delivery promises. The fix is to instruct it to mark anything it is unsure of rather than resolve it, and then to actually read the marks.

None of these are new with Flash-Lite. They exist on every model and get rarer, not absent, as models get bigger. What a downgrade changes is the frequency, which is why the rubric test above counts failures rather than asking whether the output felt good.

Is a subscription the right answer, or a different tool?

It depends which failure you hit. If the three job test shows your problem is the context window, a subscription genuinely fixes it; if it shows the problem is arithmetic or catalogue accuracy, a bigger general model is an expensive way to get a smaller version of the same error.

For the context window case, the maths is simple. Count the jobs per month that exceed the free window, estimate what each costs you in manual work, and compare with $4.99 or $19.99. Many shops will find one recurring job, usually something like a monthly review digest or a supplier reconciliation, that pays for the lower tier on its own.

For the accuracy case, a general assistant is the wrong shape. A tool that is connected to your actual catalogue, stock and order history does not have to be told what you sell, cannot hallucinate a product code that does not exist in the table it is querying, and can be checked against a record. That is a different category of product from a chat window, and the comparison worth making is not Flash-Lite against Pro but general assistant against connected tool.

There is a third option people forget: not upgrading, and splitting the job. The review digest that will not fit in 32k tokens fits comfortably if you feed it in four batches and ask for a merge at the end. It is more clicks and it is free. Whether that trade is worth it is a question about your hourly rate, and it is a legitimate answer rather than a defeat.

What this says about renting any AI tool

Capability gets re-sorted into tiers as the cost of serving it bites. That is the pattern, and it will not be the last time you see it. We have written before about what happens when the AI tool you depend on is switched off or changed under you, and the defence is the same here: keep your prompts and your inputs somewhere you own, so that moving is an afternoon rather than a rebuild.

Two things that look contradictory are both true at once. Wholesale prices for model access have fallen steeply, something we traced in detail when the labs halved their list prices without most shops seeing a smaller invoice. At the same time, the consumer app tiers are getting tighter. Those are different markets. The per token price of calling a model through an API and the price of a seat in a polished consumer app with files, memory and image generation attached are set by different pressures, and only one of them is falling.

For a merchant the practical consequence is that the answer to "should I pay for this" is not a general question about AI pricing. It is a question about which of those two markets your work belongs in. Bulk, repetitive, structured jobs tend to be cheaper through an API or a tool built on one. Occasional, varied, document heavy jobs are what the consumer subscription is good at.

What to do before Thursday

Run the three job test this week while the old behaviour is still live. That is the only part of this with a deadline, because after 9 October you cannot generate the before half of the comparison.

Then work out what the gap is worth. We collected what owners actually report spending in our breakdown of monthly AI costs for a small business, and the honest finding there was that most of the spend goes on tools nobody evaluated after the first month. A tier change is a reasonable prompt to do that evaluation.

Check whether you are paying twice. Plenty of shop software now ships an assistant inside the product you already pay for, and it may be running on a model that is perfectly adequate for the job you were taking to a separate chat window. If the AI you need is mostly attached to your catalogue and your orders, the tool that already holds that data has an advantage a general assistant does not, which is one reason our own plans and credit pricing are built around the work rather than around a seat.

Finally, write down which model you are using and when you last checked. Nobody does this, and it is the reason a change like 9 October arrives as a surprise rather than as a diary entry. A line in a notebook saying which assistant does which job, reviewed once a quarter, is enough to turn vendor churn into maintenance.

The candle maker will probably be fine. Most of her Thursday is summarising and rewriting, which is the work Flash-Lite was built for. The supplier price list is the one that will break, and she will find out in the worst way, by trusting an answer about rows the model never saw. Knowing which of your tasks is the price list is the whole job this week.