AI in my small business: test Gemini 4 Argon by 1 December

To test Gemini 4 Argon in a small business, run one real job you already did by hand, log what each run costs in rands, and decide before 1 December. Argon costs $2 per million input tokens and $10 per million output now. On 1 December that doubles to $4 and $20. On our sizing of one email job a week, a month of testing costs well under R100.

Google released Argon this week through the Gemini API and to Gemini app subscribers. Simon Willison's launch notes, dated 13 October, give the price and the date it doubles; TechCentral dates the release to 14 October and says it came after months of delays. Google's own Gemini API pricing page confirms the launch price runs to 30 November, and that the output price includes the model's thinking tokens. At about R16.50 to the dollar in mid-October, the API costs about R33 per million input tokens and R165 per million output, rising to R66 and R330. This guide covers how to use AI in my small business for one job, priced, with a date to decide.

Before you start

• One job you did by hand last month, with the finished answer kept. A month of supplier emails and the reconciliation list you made from them is a good one.
• A Google account with billing set up for the Gemini API. Argon is not on the API's free tier, and South Africa is on Google's list of supported countries.
• A spreadsheet with five columns: date, job, tokens in, tokens out, rand cost.

Step 1: choose a job with a known right answer

A test is only useful if you can mark it. Pick work you have already finished, so you can compare the model's answer with your own line by line. Avoid anything where you would not notice a mistake.

Step 2: choose the API or a subscription

The sources disagree on which subscription carries Argon. Willison calls it Google's "$100/month subscription plan". Google's South African subscription page points Argon users to Google AI Pro at R429.99 a month or AI Ultra from R1,859.99, and TechCentral also names Pro and Ultra. A subscription gives you Argon in the Gemini app, where you paste each job in by hand and see no token count.

The API charges per use and reports the tokens for every run, which is what this test needs. Our assumption for the email job: 200 emails at 1,500 tokens each is 300 000 tokens in, and a 20 000-token list out. That run costs $0.80, about R13 today and R26 after 1 December.

Step 3: strip the personal information

Before anything leaves your machine, remove ID numbers, bank account numbers and home addresses from the emails. A supplier's name and an invoice amount are usually enough for a reconciliation. What you never send cannot leak.

Step 4: run it once at each thinking level

Google calls the setting the thinking level, and Argon takes low, medium and high, with medium the default. Willison ran it at all three. His most expensive single drawing, at high, cost 17.166 US cents, about R2.83. Run your job at low first. Only move up if the answer is wrong, because a higher level uses more tokens on the same job, and thinking tokens are billed as output.

Step 5: log the cost of every run

After each run, note the input, output and thinking tokens from the usage figures the API returns with the response, and work out the rand cost in your spreadsheet. Do not estimate. The point of the test is a real number you can hold the new price against.

Step 6: mark the answer against yours

Go through the model's reconciliation list line by line against the one you made. Count three things: lines it got right, lines it got wrong, and lines it missed. A list that is 95% right but misses the one disputed invoice is not ready to run unchecked.

Step 7: repeat weekly until the end of November

One run tells you little. Run the same kind of job once a week, on that week's emails, and log each one. Four or five runs show whether the cost and the error count are stable.

Step 8: decide on 30 November

Put a reminder in the diary for 30 November, the last day of the launch price. Multiply your average run cost by two, then by the number of runs you would do in a month. If the job is worth that, keep it. If not, stop before the new price applies.

What breaks here

The rand. The price is in dollars and so is your card charge. At about R16.50 the numbers above hold. If the rand weakens, every rand figure in your spreadsheet moves with it, so log the exchange rate on each line.

POPIA. Sending customer or staff personal information to a model hosted abroad is a transfer outside South Africa. Section 72 of POPIA allows that only on set grounds, such as the person's consent or a recipient bound by adequate data protection. Step 3 exists for that reason. If the job cannot be done without personal information, it is the wrong first job.

Benchmarks are not your job. Willison notes that price per token says little on its own: what matters is how many tokens a model uses to finish a task. Artificial Analysis spent $2,407 running its benchmark on Argon at high, which Willison puts well below the latest models from OpenAI, Anthropic and Meta. But your emails are not that benchmark. Your log is the only figure that applies.

When this stops being enough

One job, one model and one spreadsheet is the right size for a test. It stops being enough when the reconciliation list has to go into the accounts, the emails arrive in three inboxes, and a person still re-types the result into Xero.

Zorah connects the AI step to the systems a business already runs. The model reads from the inbox, its answer lands in the accounts for a person to approve, and each run is logged as part of the automation itself. On Monday, pick last month's job and open the spreadsheet. The first run takes an hour.