When a new language model launches, the headline is usually about what it can do. The quieter story is what it costs to use. AI model pricing shapes which products get built, which ones survive, and how much you pay for features that now show up inside everyday apps. If you’ve ever looked at a pricing page and wondered what a ‘million tokens’ is worth, this one is for you.
Tokens are the unit of account
Most AI providers don’t charge per question or per hour. They charge per token, which is a chunk of text roughly a few characters long. A short common word might be one token, while a long or unusual word gets split into several. Both the text you send in and the text the model sends back are counted, and output tokens usually cost more than input tokens.
That has a few practical consequences:
- Long prompts cost more than short ones, even if the answer is brief.
- Chatty answers cost more than concise ones.
- Pasting a whole document into every request adds up fast.
- Conversation history is often resent with each turn, so long chats get pricier as they grow.
Why the same model family has several prices
Providers typically sell a ladder of models. The largest and most capable one costs the most, and smaller, faster versions cost a fraction of that. The reasoning is simple. Bigger models need more computing power per token, and that power is expensive. Many everyday tasks, like sorting emails or extracting names from text, don’t need the biggest model at all.
A sensible habit is to start with the cheapest model that might work, test it on real examples, and only move up if quality falls short. Plenty of teams do the opposite, pick the top model by default and never revisit it.
What pushes prices down
Over time, the cost of getting a given level of quality has tended to fall. A few forces are behind that, and none of them are mysterious.
Hardware and efficiency
Chips improve, and so does the software that runs models on them. Engineers keep finding ways to serve the same model with less memory and less waste. Savings from that can be passed on, or at least used to defend a price.
Competition
When several providers offer something similar, a lower price is one of the few clear levers left. If one company cuts rates, others tend to respond, or they point to extra features to justify staying higher. Customers who can switch easily have real bargaining power here.
Smaller models getting better
Compact models keep improving, so the ‘good enough’ tier gets cheaper while the top tier stays pricey. That widens the range of tasks where AI is affordable to run at scale.
Price wars aren’t free lunches
Low prices are good for builders, but they’re not always sustainable. Serving a model costs real money, and a provider might price below cost for a while to win market share, perhaps funded by investors. That’s a strategy, not a law of nature. If you build a product that only works at today’s rate, you’re taking on a risk that the rate changes.
There’s also the question of what you give up. A cheaper plan may come with tighter rate limits, slower responses at busy times, or fewer guarantees on uptime. Always read the limits as carefully as the price.
How to keep your own bill under control
If you’re building with an AI model, or just choosing between tools that use one, a few habits pay off quickly:
- Measure before you scale. Log token counts for a sample of real requests and multiply out to your expected volume.
- Trim prompts. Remove repeated instructions and unneeded context. Short, clear prompts are cheaper and often work better.
- Cap output length. Set a sensible maximum so one runaway answer doesn’t cost you.
- Cache what repeats. If the same question comes up often, store the answer instead of paying again.
- Avoid lock-in. Write your code so swapping one model for another is a small change, not a rewrite.
- Set spending alerts. A loop bug at 3 a.m. can be a very expensive surprise.
The bigger picture
Falling prices change what’s worth building. A task that was too costly to automate a while ago, like summarizing every support ticket or checking every document, can start to make financial sense. At the same time, cheap doesn’t mean free, and usage tends to grow to fill whatever budget exists.
The takeaway is to treat model cost as a design input from the start, the same way you’d think about storage or bandwidth. Know your tokens, match the model to the job, and keep the door open to switching. Then when the next price change lands, you’ll read it as an opportunity instead of a scramble.
