AI Costs: What the Price Drops Really Show
Yes, many AI tasks have become much cheaper per unit of output, especially when measured by the price of processing text through a model API. But there is no single “AI cost” figure: model prices, the cost of running a service, and what users ultimately pay are different things, and headline drops often compare different models or capabilities.
What has actually become cheaper?
The clearest price comparisons come from providers’ published rates for API use, usually quoted per million input and output tokens. A token is a short piece of text; it is not exactly a word, and the token count depends on the language and content. Input tokens are what a model reads, while output tokens are what it generates. Providers may charge different rates for each.
Published prices have fallen sharply for some tasks as newer, smaller, or more efficient models have appeared. Stanford’s 2025 AI Index, for example, reported that the inference cost for a system performing at a benchmark level comparable to GPT-3.5 fell by more than 280-fold between November 2022 and October 2024. That comparison is tied to a specific performance benchmark and a selected set of model prices; it does not mean every AI task, product, or bill became 280 times cheaper.
The unit matters. A lower cost per token does not guarantee a lower total cost if a product uses more tokens, handles more requests, or adds costly tools such as image generation and web search. Nor does a lower API price necessarily mean a consumer subscription or business contract falls by the same proportion.
Why the numbers fall
Several changes can reduce the cost of producing a given response. Hardware has improved, and software can use computing resources more efficiently. Researchers and developers have also built models that deliver useful performance with fewer parameters or less computation for particular tasks. Competition and new service offerings put further pressure on published rates.
Serving a model also involves choices about how computing capacity is used. Providers can batch requests, route straightforward questions to smaller models, or cache repeated results. These techniques can reduce the resources needed per request, although availability and savings vary by service and workload.
- More efficient models: A smaller model may handle routine classification or summarization at lower cost than a large general-purpose model.
- Better hardware and serving: Faster computation and improved scheduling can lower the cost of generating each token.
- Different price tiers: A provider may offer cheaper models with fewer capabilities, alongside more expensive models for complex work.
- More competition: Providers may cut prices or introduce low-cost options to attract customers.
Why one cost-drop statistic can mislead
“AI got cheaper” can refer to at least three different measures: the price a provider charges, the provider’s cost to serve a request, or the total cost to a user or organization. Public price lists usually reveal the first, not the second. The actual cost of building and operating a model can include research, training, data centers, energy, staff, and infrastructure, and those costs are not captured by an API rate.
Comparisons also depend on what counts as equal performance. A cheap model may be suitable for simple questions but less reliable on long, difficult, or specialized tasks. A benchmark can give a useful common test, but it cannot represent every real-world use. A price-per-token comparison between two different models is not an apples-to-apples measure of the cost of producing an equally good answer.
Finally, token prices are not the whole bill. Some services charge separately for images, audio, search, or other tools; some have minimums, usage tiers, or contract terms. A user interface may include hosting, support, storage, and product features rather than simply passing through API charges.
What lower prices mean for real users
For an individual using a free chatbot, a drop in underlying model costs may show up as better access, more generous limits, or improved capability rather than a lower subscription fee. A paid plan is priced as a product, not as a direct meter of the model’s compute cost.
For a business using an API, estimate the bill from the actual workload. Suppose a task uses 1,000 input tokens and generates 500 output tokens. Multiply each amount by the applicable per-token rate, often listed per million tokens, and add any charges for tools or other services. Then multiply by the expected number of tasks. The result is still an estimate: prompts can vary in length, retries add usage, and some systems include conversation history in each request.
Lower unit prices can also increase use. If each request becomes cheaper, a company may automate more tasks or ask models to produce longer responses. That can make the total bill stay flat or rise even while the cost per token falls—a common effect when cheaper services encourage greater consumption.
How to compare AI prices fairly
Before choosing a model or service, compare the cost of completing the task to an acceptable standard, not just the cheapest listed token rate. Check the current pricing page because rates and product terms can change.
- Use the same task, prompt, and output-length assumptions for each option.
- Separate input and output charges, and include tool or media fees.
- Test quality on representative examples, including difficult cases and likely failure modes.
- Measure tokens and retries on a sample workload, then project monthly usage.
- Account for latency, reliability, privacy requirements, and the cost of human review.
A low-cost model that needs frequent correction may be more expensive in practice than a pricier model that completes the task reliably. For routine work, a smaller model may be entirely sufficient. For high-stakes decisions, cost alone is not an adequate selection criterion.
Frequently Asked Questions
Is AI getting cheaper?
Many AI services have become cheaper per unit of model output, particularly for API inference. The trend is not uniform across models, features, or products, and lower unit prices do not ensure a lower total bill.
What does “cost per token” mean?
It is a charge based on the amount of text a model processes or generates. Input and output tokens may have different rates. Tokens are fragments of text, so one token does not always equal one word.
Does a 280-fold cost drop mean AI is 280 times cheaper for everyone?
No. That figure describes a particular benchmark-based comparison of inference costs over a stated period. It does not describe every model, task, subscription, or organization’s total expenses.
Why hasn’t my AI subscription gotten cheaper?
A subscription price covers a product and its features, not just the underlying model’s per-token compute. Providers may use lower costs to add capability or capacity rather than reduce the monthly fee.
Can cheaper AI increase my total costs?
Yes. If lower prices lead to more requests, longer outputs, or additional tool use, total spending can rise even as the price for each unit falls.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.