What's the catch with cheap AI? A scope limit

What's the catch with cheap AI?

The catch with cheap AI is scope. A $5 app runs a model a step behind the frontier and leaves out the expensive extras: image generation, project workspaces, building you an app. The low price comes from cheaper models, and for everyday questions the answers hold up.

My answer for Plainly is that the catch is a scope limit. It runs a value model, a step off the frontier, and it's ruthlessly optimized for chat. It is very good and quick at answering questions. It won't make you an image, it won't give you twenty projects in a big desktop interface, and it won't build you an app. For a specific kind of person, that is all they need.

People who ask "what's the catch" usually mean four more things, so here are short answers before the detail.

Why cheap AI can be cheap

Cheap AI runs on value models: open-weights models from labs such as Z.ai and DeepSeek that sit a step behind the frontier. Open weights mean any host can download and serve them, so hosts compete on price. Closed frontier models are sold only by the lab that made them, at the lab's price.

Sources: OpenRouter model list and claude.com/pricing, 11 October 2026; Arena text leaderboard, updated 8 October 2026 (scores for each model's highest reasoning setting where Arena lists one).

Kimi K3 shows that open weights alone don't make a model cheap. It scores higher than the two Flash models and charges eight to twenty times more per output token, closer to a closed frontier model's pricing than to the value tier.

My own bill is one data point. I'm a very heavy AI user. Through a desktop chat app running my own OpenRouter key (a setup for technical people), the two value models I use for everyday chat, DeepSeek V4 Flash and GLM-5.3 Flash, cost me $18.90 between them for the month to 15 September 2026. Those were my personal API bills, separate from Plainly.

What the scope limit leaves out

Plainly can explain how to build an app and write code for it, but it won't build and run the app for you. It can describe a scene in words and can't draw one.

How big is the quality gap?

For everyday chat, small. On Arena's text leaderboard, where people vote between two anonymous answers, GLM-5.3 Flash and DeepSeek V4.1 Flash both score 1475 and Claude Fable 5 scores 1504 (8 October 2026). On Arena's rating scale, a 29-point gap means voters would pick the Fable 5 answer about 54% of the time and the value model's about 46%. The top of the board, Gemini 4 Argon, scored a preliminary 1525.

In my own writing test on 2 September 2026, five everyday writing tasks scored out of 75, ChatGPT on the Go plan and Plainly both scored 61 and Fable 5 scored 59. ChatGPT won the text-message and review tasks, Plainly won the email and the question about how a $5 app can charge so little, and the condensing task was a tie between those two. The order wasn't randomized. Every answer is printed in the best AI for writing test.

Arena's text score measures which answer people prefer in everyday chat. It says less about the hardest work: long chains of reasoning, difficult code, and specialist answers that must be right the first time. If that describes your working day, test a frontier model on your own problems, since that work is where a frontier model earns its price.

Is cheap AI safe with my data?

Price tells you little about privacy. The provider's policy tells you what happens to your chats. Read it for three things: whether chats train models, where they are stored, and who at the company can read them.

Some of the cheapest options do raise real concerns. DeepSeek's own app stores data in China, uses it to train models, and takes training opt-outs by email (DeepSeek privacy policy, updated 10 February 2026). The same open-weights model served by a different host follows that host's policy instead. Paid plans are no guarantee either. ChatGPT trains on consumer chats unless you opt out, on every consumer plan, and in 2025 Anthropic moved Claude Free, Pro and Max to training unless you opt out, with five-year retention for anyone who leaves it on (TechCrunch, 28 August 2025).

On Plainly, chats are never used to train AI. We don't read your chats unless you ask us to or the law requires it. Plainly's OpenRouter account excludes providers that train on prompts, though some hosts keep prompts for a period. Does Plainly train on your chats shows where a message goes, and are Chinese AI models safe covers the China question.

Will cheap AI go up in price?

Frontier AI is subsidized, so its price could rise. OpenAI took in $13.1 billion in 2025 and still lost money, and in January 2025 Sam Altman said the company was losing money on its $200 Pro plan because "people use it much more than we expected" (TechCrunch, 5 January 2025). A subsidized price can rise later, or the plan can stay the same price and give you less.

Value models sit on a different curve. Epoch AI found in March 2025 that the price of reaching a fixed level of AI performance was falling between 9 and 900 times a year, depending on the task. An app built on value models gains when a better one ships at a similar price. Nobody can promise a price forever, and an app that bills monthly and lets you cancel in settings lets you leave the day it changes. The future of AI pricing sets out where prices are heading.

Is cheap AI a scam or trial bait?

Five checks catch trial bait and headline prices that mislead:

1. Find the monthly price. Claude Pro's $17 a month, for example, applies only when you pay $200 for a year up front; monthly billing is $20.
2. Look for credits. A plan that sells a pot of credits charges you per message, however it is labeled.
3. Check what happens after an introductory price. A first month at $1 that renews at $30 is a $30 plan.
4. Find the cancel button before you pay. It should be in settings, with no email or phone call needed.
5. Be wary of "lifetime" deals for frontier models. Each answer costs the seller money every time, and a one-off payment can't cover that forever.

Plainly is $5 a month, billed monthly, with no credits, and you cancel any time in settings. New accounts can ask a limited amount free each month before deciding. How the $5 pays for itself is covered in how a $5 AI app makes money.

When to pay more than $5

The scope limit matters if your work runs on the hardest problems in your field, on images or video, on several projects at once, or on shipping software. Then a frontier subscription covers you and a $5 app won't grow into it.

My rule of thumb: if you use AI a little, the $20 Claude plan; if you use it a lot, the $5 Plainly plan; if you use it a huge amount, the $100 Claude or ChatGPT plan.