Not long ago, the cost of using the AI model itself could have a major impact on the economics of an AI project. That balance is now changing quickly, as newer models become more capable and, for many tasks, significantly cheaper to use.
For companies, this means more than just a lower API bill. Falling prices can change the economics of projects that did not make sense a year ago, especially when AI handles a large volume of repetitive tasks or represents a significant share of operating costs.
In this article, we look at how quickly the cost of AI performance is falling, which business processes may benefit the most, and why model pricing alone is still not enough to determine whether an AI project makes financial sense.
The same level of AI performance costs significantly less than a few years ago
A September 2026 analysis by research organization Epoch AI: The plunging price of thought shows the extent of the change.
According to the analysis, the cost of reaching a comparable level of performance across five AI benchmarks has fallen by an average of approximately 47% per quarter since 2023. That translates to roughly a thirteenfold decline per year.
One example shows just how significant the shift can be.
At the beginning of 2025, OpenAI’s o3 model reached a certain level of performance on the demanding GPQA Diamond benchmark at an estimated cost of around $0.30 per question. Less than 18 months later, GPT-5.6 Luna achieved a comparable result at approximately $0.0004 per question.
The cost of reaching a comparable result fell by roughly 725 times.
A benchmark is not the same as a real business process, of course, and the same rate of decline cannot simply be applied to every type of AI solution. The authors themselves point out this limitation.
The data does show, however, that the same budget can buy significantly more model performance today than it could just a few years ago.
For some companies, that is a reason to revisit projects that previously did not make economic sense.
Projects that were too expensive a year ago may be worth recalculating today
A use case that was too expensive last year does not have to remain a bad idea forever. If model usage represented a significant share of its operating costs, current prices may change the calculation.
The biggest potential is in processes where the same operation is repeated frequently, where costs increased rapidly as the solution grew, or where an expensive, high-performance model had to be used for a large number of relatively simple tasks.
Examples include document processing and classification, extracting information from text and databases, data validation and categorization, customer requests, recurring reports, or automations connected to CRM, ERP and other business systems.
For a single operation, a difference of a few cents may be negligible. But if the same process runs tens of thousands or millions of AI operations every month, even a small difference starts to affect total operating costs.
That does not mean every previously rejected AI project should be reopened. Recalculation makes the most sense where model pricing or the cost of scaling was one of the main reasons the solution did not move forward.
September’s new models showed that prices are still falling in 2026
The longer-term trend was also visible in September 2026, when OpenAI and Anthropic introduced new models almost on the same day.
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. When launching the model, OpenAI said this represented a 50% reduction compared with the then-discounted prices of GPT-5.6 Sol.
For the lower-cost GPT-6 Luna, one million input tokens costs $0.10 and one million output tokens $0.50. Here too, prices are significantly lower than for the previous generation.
Anthropic moved in a similar direction with Claude Opus 5.5. Its list price is $4 per million input tokens and $20 per million output tokens. The token price itself is lower than with Opus 5, while Anthropic says that typical workloads cost approximately 40% less overall.
A summary of the September changes and pricing is also available in CzechCrunch’s article on the new OpenAI and Anthropic models.
But the price list still does not tell you how much it will cost to complete a specific task or an entire process. Newer models may complete the same work in a different number of steps, with a different success rate and different requirements for human review.
A cheaper model does not necessarily mean a cheaper process
Consider two models.
The first costs less per call, but has to repeat some tasks, produces incorrect outputs more often and requires more human review. The second is more expensive per call but completes a larger share of the work correctly on the first attempt.
A lower token price does not necessarily mean a lower cost per result.
For business deployments, it is therefore more useful to track metrics such as the cost of one correctly processed document, one resolved customer request, one completed report or another specific output.
With more complex automations and AI agents, this difference becomes even more important. A single task may require retrieving information, working across several systems, evaluating data, performing an action and then checking the result.
One request can therefore trigger several or even dozens of model calls.
As the number of steps increases, what matters is not only the price of each call, but also how many steps the model needs, how many it completes correctly and how much work is still left for a person.
The biggest impact may be in high-volume processes
Falling prices matter most where companies use AI repeatedly and across large volumes of work.
If AI processes thousands of documents, categorizes customer requests, checks data or works with a large internal knowledge base, even a relatively small difference in the cost of one operation can add up over time.
At the same time, there is more room to combine different models.
Simple and clearly defined tasks can be handled by a lower-cost model, while more complex or higher-risk cases can be routed to a more capable model or to a person.
Basic document classification, for example, does not require the same model capability as a complex contract analysis. Likewise, a routine customer request does not need to follow the same path as an unusual case with a high cost of error.
Instead of searching for one universally “best” model, it may be more effective to split a process according to the required quality, cost and level of risk.
The model is still only one item on the total bill
A significant drop in model prices does not mean the entire AI solution becomes cheaper at the same rate.
Companies still need to deal with integrations into existing systems, data availability and quality, infrastructure, security, monitoring, human oversight and ongoing maintenance.
In more complex projects, these areas can represent a substantial part of the overall budget.
If an AI automation needs to work with a CRM, accounting software and an internal database, the model itself represents only one part of the economics of the solution.
We looked at the full cost structure in more detail in our article How much will AI cost in 2027? The model price is only part of the bill.
A new API price therefore cannot simply be applied to the original project budget with the expectation that the entire solution will become cheaper by the same percentage, the whole process needs to be recalculated.
Price only matters in the context of a specific process
Even much cheaper AI will not turn a poorly chosen process into a good project.
If an automation saves a few seconds on a task that a company performs ten times a month, a drop in model pricing will probably not make a meaningful difference.
The situation is very different for a process that handles tens of thousands of requests every month or consumes a significant share of a team’s capacity.
Before selecting a particular model, a company therefore needs to understand:
- how much work actually flows through the process
- how much it costs today
- how much employee time it consumes
- where errors occur
- how much can realistically be automated
- how much human oversight will still be required after AI is introduced.
Only then can the company determine how much the falling cost of the model can actually affect the return on the overall project.
What does this mean for companies?
The decline in the cost of AI performance over the past few years has been significant enough that some calculations made in 2025 may no longer reflect the conditions at the end of 2026.
Companies can therefore revisit use cases that were previously rejected because of operating costs and recalculate them based on current model prices, capabilities and actual workload volumes.
For some projects, almost nothing will change because integrations, data or human work still make up most of the cost. For others, a cheaper and more efficient model can significantly reduce the cost per completed task and change the economics of the entire solution.
What matters is not simply how much a particular model has become cheaper, but whether that change has altered the economics of the business process the company actually needs to solve.