ChatGPT, Claude or Gemini for a small business: what changed this quarter
Two vendors shipped new models in the first two weeks of September. None of it changes the answer for a founder with one laptop and no IT department. Pick the problem first, then pick the tool that reads your own files best. Here is what actually moved, and what it means at your desk.
As of 13 September 2026.
What changed in the last two weeks?
Two release days, three new model names, no change to the shape of the decision. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. Mythos 5.1 is the same underlying model at a different safeguard tier, available only through restricted-access programs for vetted cybersecurity and life-sciences organizations, so you cannot buy it and do not need it (anthropic.com). Google released Gemini 3.8 Flash on 2 September, alongside a Flash Cyber variant aimed at finding software vulnerabilities. The gains Google reports for 3.8 Flash are in long-horizon software engineering, multi-step reasoning, and calling tools repeatedly on a hard task (blog.google). Separately, with no September date on it, the model OpenAI now points developers at is GPT-6 Astra (developers.openai.com).
Read that list again. Every line of difference is about coding, security research and long technical work. Not one of them is about the quotation you have to send on Monday.
What AI tools should I use for my business?
One, and the one you already have open. You do marketing, research and finance yourself. You do not have the hours to run three subscriptions in parallel and compare them honestly. Pick the tool that handles the material you actually hold, then stay with it for a quarter.
The material decides. If your week is documents and numbers in Google Sheets and Gmail, Gemini has the shorter path. Deep Research in Gemini can draw on context from Gmail, Drive and Chat, including Docs, Slides, Sheets and PDFs, and work that into a report (blog.google). If your week is long documents and careful written argument, Claude's newest pair is built for knowledge work. If your week is mixed and you want one tool that browses, reads files and runs code in the same thread, ChatGPT covers the widest spread.
This is worth saying plainly, because the general answer you get online is the opposite. Ask that question online and a starter stack of tools comes back, followed by a warning not to buy a dozen subscriptions.
How do I use ChatGPT for work when the releases keep coming?
Put one real document in front of it and check the output against something you already know. That is the whole method, and it survives every release. I learned AI by asking one question of my own work each day, not by reading release notes.
A worked version. Take your last three months of sales data and ask for the top five insights, the questions a buyer would ask, and the next actions. Then check the numbers against your own ledger. Where the output is right, you have saved an afternoon. Where it is wrong, you have learned where this tool is weak on your kind of data, which is the more useful finding.
One detail does matter on a free account. The default model for ChatGPT Free and Go accounts is GPT-5.6 Luna, which OpenAI describes as designed for cost-sensitive, high-volume workloads (developers.openai.com). That is the fast one, not the careful one. If the work is analysis you will sign your name to, use a paid account and the better model, or do the analysis in two passes and check both.
What to ignore this quarter
Benchmark numbers. A top score on a reasoning test tells you nothing about whether a model reads your supplier contracts correctly. The same goes for every percentage in every launch post. None of those tests used your documents.
Ignore the new-model anxiety too. A prompt that works on a real task keeps working. What changes is speed and cost, and those only matter once you have a task that runs often enough to count.
Ignore, finally, the advice to wait for things to settle. The skill you are building is the judgment to tell whether an output holds. That skill transfers to whatever ships next quarter.
What else would you ask before switching tools?
Should I pay for two tools instead of one?
Not in the first quarter. Two subscriptions means two sets of habits and no clear read on either. Run one tool on one recurring task until you can predict where it fails. Then, if a second tool clearly covers that gap, add it.
Does a new model break the prompts I already wrote?
No, in the ordinary case. A brief that states the role, the context, the task, the format and the constraints keeps working across releases, because it describes your work and not the model. Re-check the output once after a release, on a task whose correct answer you know.
Which tool is best for research on my market?
Whichever one you make show its work. Ask for a source URL against every claim in the output, then open two of them yourself. If a page does not exist, or does not say what the output said it says, treat the rest of that answer as unchecked and go back for the sources again.