What the newest AI models change for a small business

Written 8 September 2026. That date matters more on this page than on anything else we publish, because this field moves faster than we can keep a page current. Treat what follows as a snapshot with its sources attached, not as a page that will still be accurate in six months — and there is a section at the end on how to check what is true when you read it.

Three months, three frontier labs, and a great deal of noise. Here is what actually shipped, and then the more useful question: which of it changes anything for a business with fifteen staff and a website.

What shipped

OpenAI — GPT-5.6 in July, GPT-6 Astra on 3 September. OpenAI describes Astra as improved at coding, research, computer use and multi-step work, and says it can produce documents, spreadsheets and presentations that follow your own templates and adapt when you change the requirements mid-task. It rolled out to ChatGPT Plus, Pro, Business and Enterprise, and through the API, Azure and AWS Bedrock.

The more interesting document is the system card, which is unusually candid in two directions at once. It calls Astra the most capable model OpenAI has broadly deployed and says it “can find previously unknown security flaws and develop new ways to exploit them” — and it also records that Astra shows decreased chain-of-thought monitorability than earlier models, which is OpenAI saying, in its own safety documentation, that its newest model is harder to watch think. Both of those tell you more than any benchmark on the announcement page.

Anthropic — Claude Opus 5 in July, Fable 5.1 and Mythos 5.1 in August. Opus 5 was framed around long-running agents; the August pair is described as their most advanced for coding and knowledge work. Anthropic also published how Claude’s text watermark works in mid-August, which is a notable thing for a lab to document rather than leave implicit.

Google — three Gemini models in July, then Gemini 3.8 Flash and 3.8 Flash Cyber. Google’s own framing is worth quoting almost directly: better reasoning and coding at the same speed and cost as 3.7, and it is the third Flash release in six weeks. Pricing held at $0.75 per million input tokens and $3.75 per million output, with that introductory rate stated to expire on 31 December 2026.

You will notice we have quoted almost no benchmark scores. That is deliberate. Every one of those numbers is the vendor’s own, measured on tests the vendor chose, and none of them predicts whether the thing will write a decent product description for your shop.

The pattern that matters more than any single release

Three Flash releases in six weeks. That is the headline, and it is not about Google.

For a small business it means the question “which AI is best?” has stopped being answerable in a way that stays answered. Any comparison you read is describing a leaderboard that will have changed by the time you have finished acting on it. Building a process that depends on one specific model being the best one is building on something designed to be replaced next month.

The practical consequences:

• Do not sign a long contract on the basis of one model’s benchmark. By renewal it will be two generations old.
• Keep your prompts and your data separate from the vendor. If your workflows live inside one tool’s proprietary format, switching costs you the workflows.
• Pick on price, latency and whether it does your actual job, not on which is nominally smartest. For most business tasks — writing a description, summarising a call, sorting enquiries — the gap between the frontier model and the cheap fast one is smaller than the gap between a good prompt and a lazy one.

The one capability that genuinely is new

Not the intelligence. The agency — models that use a computer, browse, and carry a multi-step task through to the end rather than answering one question at a time. It is the common thread across all three labs this quarter: OpenAI’s framing of Astra around computer use and multi-step work, Anthropic’s around long-running agents, Google’s around long-horizon software engineering.

For a small business, that is the difference between “it writes a draft for me” and “it does the task”. Reconciling a statement, pulling forty invoices into a spreadsheet, going through a mailbox and sorting it. Those are real jobs and they are becoming possible.

The honest caveat: an agent that can complete a task can also complete it wrongly, at speed, and without stopping to ask. Give one write access to something that matters and you have created a new category of Monday morning. Start it somewhere reversible.

The thing nobody mentions: they do not know what happened recently

Google’s own model card puts Gemini 3.8 Flash’s knowledge cutoff at March 2026. Every one of these models has a date past which it simply does not know anything, and it is usually months before you are talking to it.

This catches businesses constantly. Ask about a tax rule, a regulation, a competitor’s current pricing, and you may get a confident answer describing a world that has moved on. The models that can browse will go and look; the ones that cannot will answer from memory and sound identical either way.

If the answer depends on a fact that could have changed, make it search, and check the link it gives you. That is not a limitation of one model. It is what these things are.

What we would actually do with this, in a business your size

1. Pick the cheap fast model first. Try the task on the least expensive option that could plausibly do it. Most business writing and summarising does not need the frontier.
2. Write the process down, not the tool. “Draft a description from these five fields, in this tone, under 80 words” survives a model change. A saved chat does not.
3. Use it where being wrong is cheap. First drafts, summaries, sorting. Not final numbers, not legal wording, not anything that goes out unread.
4. Keep a human on anything with your name on it. Not for principle — because the failure mode is confident and fluent, which is exactly the kind that gets through.
5. Revisit in three months. Given the release cadence above, anything you conclude today has a short shelf life. That is annoying and it is the actual situation.

How to check whether this page is still true

Since it will go out of date, here is how to replace it rather than trust it:

• The labs publish their own announcements. OpenAI, Anthropic, Google. Those pages are the primary source; every article about a release, including this one, is downstream of them.
• Model cards carry the details that press coverage drops — knowledge cutoff, context limits, intended use.
• Be sceptical of “latest AI news” roundups, this one included. While researching this piece, several aggregator sites confidently placed Anthropic’s August releases in September. We only caught it because we checked the announcement page. If a page does not link to a primary source, treat it as someone’s guess.

---

We build the automations that sit on top of this — the connection between a model and your actual systems, which is where the value is and where the work is. If you have a task in mind and want to know whether it is worth doing yet, describe it to us. Sometimes the answer is not yet, and we would rather say so.

Sources: GPT-6 Astra system card · GPT-5.6 launch · Anthropic news · Gemini 3.8 Flash and 3.8 Flash Cyber · Gemini 3.8 Flash model card