Google is at it again. Since 2 September 2026 there has been Gemini 3.8 Flash, three weeks after its predecessor 3.7 — making it the third Flash model in six weeks. If this pace has made you tune out: understandable, and you are missing less than the headlines suggest. One thing you should take away, though: The real news is not what Gemini 3.8 Flash can do. It is the price it comes at.
I will sort this the way I did with the GPT-6 update from OpenAI: what is new, what it costs, who gets it — and what actually matters for you as a solopreneur running an online business.
What is Gemini 3.8 Flash?
Flash is Google’s workhorse class: the models that run fast and cheap and therefore do most of the work — unlike the big Pro models, which mostly deliver headlines. Google calls 3.8 Flash its “most intelligent workhorse”: better at coding, at AI agents and at multi-step reasoning, at the same price and speed as its predecessor.
The test results behind that are remarkable: at solving complete software tasks on its own, 3.8 Flash beats most of the much larger frontier models — at a fraction of the cost. In a test of expert questions across all fields it reaches 54.9 percent, another score that used to require the expensive models. The price for that: the model “works harder”, as Google itself puts it — more intermediate steps, more tool calls, and therefore sometimes more tokens used per task. Keep that in mind for the next section.
A side note worth noticing: Google now ships Flash models at a weekly pace — while the big Pro flagship, originally announced for June, is still nowhere to be seen.
What does Gemini 3.8 Flash cost?
In the subscription: nothing extra. Gemini 3.8 Flash is included in the paid plans Google AI Pro and Ultra — in the Gemini app, in AI Mode in Google Search and in Google Sheets. If you use Gemini for free, you do not get the new model.
It gets interesting with the usage prices — the ones automations and AI agents are billed at:
| Price per M tokens (API) | Gemini 3.8 Flash | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|
| Input | $0.75 | $10 | $10 |
| Output | $3.75 | $50 | $50 |
For perspective: the frontier models Claude Fable 5.1 and GPT-6 Astra cost a good thirteen times as much per million tokens. In some tests Gemini 3.8 Flash gets close to their level — for a thirteenth of the price. Two honest footnotes belong here: it is an introductory price that expires on 31 December 2026 — after that it doubles to 1.50 and 7.50 dollars. And because the model “works harder”, it sometimes uses more tokens per task than its predecessor. Neither changes the picture: a year ago, this class of performance cost many times more.
Why the price is the real news
In your online business there are two kinds of AI tasks. One needs the best available model: strategy, positioning, copy where every nuance counts. The other is assembly-line work: pre-sorting enquiries, pulling data out of documents, writing summaries, updating descriptions — tasks that do not need to be brilliant, but reliable and cheap, because they run a hundred times.
Exactly this second kind is currently getting so cheap that price stops being a counter-argument. An AI agent that takes over your routine work pays for itself at 0.75 dollars per million tokens as soon as it saves you a few minutes. That was not true six months ago — and it holds no matter whether that agent ends up running on Gemini, Claude or something else entirely, because the price pressure applies to every provider.
One detail from the announcement that gets lost in the headlines but matters for automations: according to Google’s measurements, 3.8 Flash is considerably more robust against prompt injection — hidden instructions in other people’s documents and websites that try to trick an AI agent into doing something other than what you told it to do. The more you let AI work on its own, the more this unglamorous property matters.
And Gemini 3.8 Flash Cyber?
The second variant in the announcement is newspaper knowledge for you: 3.8 Flash Cyber is specialised in IT security — it finds vulnerabilities and writes fixes, according to Google’s Chrome team 2.6 times more usable ones than much larger models. Access is only available through Google’s new Fairwind programme for government agencies, operators of critical infrastructure and vetted security firms.
You know the pattern by now: OpenAI ships GPT-6 only in a restricted version, Anthropic keeps Mythos for vetted organisations, Google goes straight to building its own access programme. For your day-to-day this is irrelevant — it just shows how seriously the providers now take their own models.
Do you have to switch to Gemini now?
No — and the reasoning is the same as with the two model updates of the past week: the gap between the providers is smaller than the gap between a tool that lives inside your workflows and one you would first have to get to know. Where your projects, instructions and routines live — that is where your edge lives. Which provider fundamentally fits you is something I took apart in ChatGPT Astra or Claude Fable 5.1 — Gemini 3.8 Flash changes nothing about that.
And even without switching, this model family already crosses your path: as the engine behind Google Search, behind Google Sheets and behind NotebookLM. If you work there, you benefit from every one of these updates automatically.
What you should do now
- If you use Gemini for free: nothing. You do not get 3.8 Flash — and for everyday tasks you lose little. How far you really get with free AI is something I have written up separately.
- If you have Google AI Pro or Ultra: pick 3.8 Flash in the model menu and give it a real multi-step task — something with research, a table and text in one go. That is exactly where the jump over its predecessor lies.
- Whatever provider you use: write down the three routine tasks that cost you the most time every week. At these token prices, automating pays off faster than it did three months ago — and that list is the first step, no matter which model you use to implement it later.
Model news arrives every week. The real work is turning it into workflows for your business. That is exactly what we talk about in my free community “KI — aber richtig” — the templates and workflows from my videos live there too. And if you want company instead of doing it alone: the AI Business Community offers monthly live calls and workshops for exactly that.
Frequently asked questions
What is Gemini 3.8 Flash?
Gemini 3.8 Flash is Google’s new standard model, released on 2 September 2026 — the third Flash model within six weeks. It mainly improves coding, AI agents and multi-step reasoning, and in some tests comes close to the level of much more expensive frontier models.
What does Gemini 3.8 Flash cost?
Nothing extra in the subscription: Google AI Pro and Ultra include the model in the Gemini app, in AI Mode in Google Search and in Google Sheets. Via the developer API it costs 0.75 dollars per million input tokens and 3.75 dollars for output — an introductory price that runs until 31 December 2026 and doubles after that.
Is Gemini 3.8 Flash free?
No. Gemini 3.8 Flash is only available in the paid plans Google AI Pro and Ultra. If you use Gemini for free, you keep working with the previous models — which is enough for everyday tasks in most cases.
Is Gemini 3.8 Flash better than ChatGPT or Claude?
In some tests Gemini 3.8 Flash comes close to the frontier models GPT-6 Astra and Claude Fable 5.1 — at a fraction of the price per token. For your day-to-day, though, the deciding factor is less the benchmark score and more where your projects, instructions and routines already live.
It is not the model that decides — it is your list
Gemini 3.8 Flash is no reason to overturn anything — but it is a clear signal: good AI is getting cheaper faster than most people keep track of. With every one of these updates, the threshold at which automation pays off for a small business drops. The difference is not made by whoever tests the new model first, but by whoever knows her own workflows well enough to use the falling prices. Start with the list — by then, the models will only have gotten better.
