AI got cheap while nobody was looking
Everyone is arguing about AI safety and open weights. Meanwhile the price of using these tools has quietly collapsed — and if you are still paying $20 a month out of habit, you may be paying four times more than you need to.
The AI news for the past fortnight has been almost entirely about governance. A sandbox breach, a security alliance, an open letter signed by more than a thousand researchers. Serious stories, and I have written about two of them.
Underneath all of it, something happened that affects far more people day to day: the price of using AI fell off a cliff, and hardly anyone noticed.
What the tools cost now
For two years the answer to “what does an AI assistant cost” was twenty dollars a month, everywhere, for everyone. That is no longer true.
As of the end of July 2026, the main consumer subscriptions look roughly like this:
- Google AI Plus — $4.99/month, cut from $7.99 on 8 June, with storage doubled to 400GB at the same time
- ChatGPT Go — $8/month
- The standard tier — $20/month across ChatGPT Plus, Claude Pro, Google AI Pro and Perplexity Pro. Remarkably uniform.
- Power tiers — $100 and $200 depending on vendor
The number worth staring at is the first one. Five dollars a month is less than a music subscription. It buys a capable model, video generation, and Google’s research tool NotebookLM. Two years ago that would have been an enterprise contract.
The developer side fell further
If you build anything on an API rather than a subscription, the drop is steeper still.
On 30 July, OpenAI cut the price of its two cheaper GPT-5.6 models: Luna to $0.20 per million input tokens and $1.20 output, Terra to $2 and $12. The flagship, Sol, held at $5 and $30.
DeepSeek sits lower again, with its Flash model at $0.14 input and $0.28 output per million, and a 75 percent cut on its Pro tier made permanent in May.
To put that in ordinary terms: a million tokens is roughly 750,000 words. Processing an entire long novel through a cheap model now costs less than a cup of tea.
Why this is happening
Three things at once.
The models stopped being different enough. When one assistant was clearly ahead, it could charge for that. In 2026 the top models trade places monthly and most people cannot tell them apart on everyday work. When products converge, price becomes the lever.
Switching costs collapsed. For developers especially, moving between providers is now a configuration change rather than a rewrite. No lock-in means no pricing power.
The fight moved to emerging markets first. This is the part I find most interesting, and it explains the timing. Cheap tiers appeared in India well before they reached the US — ChatGPT Go launched there at under $5 a month, Google followed with a sub-$5 plan. The logic of undercutting to capture users at scale was tested in Asia, worked, and has now crossed over.
For those of us outside the US and Europe, that is a familiar pattern running in an unfamiliar direction: for once the cheaper pricing arrived in our markets first.
What you should actually do
If you pay for AI, the honest advice is short.
Check what you are on. A lot of people signed up at $20 in 2024 and have not looked since. If your use is writing emails, summarising documents and asking questions, a $5 tier will very likely do it. That is $180 a year for capability you may not be using.
Try the cheap tier before the expensive one. The old advice was to start free and upgrade when you hit limits. There is now a genuine middle step, and it did not exist a year ago.
If you are building, price-check monthly. A model that cost $2 per million tokens in June may cost $0.20 now. If you set up a pipeline six months ago and never revisited it, you are probably overpaying by a multiple.
Do not switch for the sake of it. Learning a tool properly is worth more than a few dollars a month. The saving is real; the time you lose relearning your workflow is also real.
The obvious question
Is any of this sustainable?
Probably not at these levels. Running these models costs real money in electricity and hardware, and the companies competing hardest on price are largely the ones with other businesses to fund it. Cheap subscriptions are a customer-acquisition strategy, and customer-acquisition strategies end.
What I would expect: prices stay low until the market settles, then quietly rise for anyone who stopped paying attention. Which is exactly how every subscription market has worked.
So the practical move is to enjoy it, avoid annual plans while pricing is this volatile, and check your bill every few months.
Why this matters more than it sounds
A five dollar entry point changes who gets to use this.
At $20 a month, an AI assistant is a professional tool for people in wealthy countries with employers who reimburse. At $5 it is affordable to students, to people in markets where twenty dollars is a meaningful sum, to anyone curious rather than committed.
I built my first useful automation — a macro that replaced several hours of month-end work — using a tool I was paying for anyway. If I had needed to justify twenty dollars a month up front for something I did not yet know would work, I am not certain I would have started.
That is the part the governance headlines miss. While the industry argues about who governs frontier models, the thing quietly reshaping who actually uses AI is a number on a pricing page.
Prices in this article were accurate at the end of July 2026 and are changing constantly. Check the vendor’s pricing page before deciding anything.
Related reading
