On paper, the cost of using today\u2019s generative models is falling based on what AI companies are charging.<\/p>\n
For example, OpenAI\u2019s GPT-4 with an 8K context window had cost<\/a> $30 per million input tokens and $60 per million output tokens as of early 2023. This year, GPT-4 Turbo, which is more powerful, nonetheless costs<\/a> 50% to 67% less: $10 per million input tokens and $30 for the output.<\/p>\n According to Stanford\u2019s 2025 Artificial Intelligence Index report<\/a>, as AI models become more capable and smaller, the costs for applying them in use cases \u2014 inference \u2014 \u201chave fallen anywhere from nine to 900 times per year,\u201d the report said.<\/p>\n When it comes to infrastructure, costs have declined by 30% annually, while energy efficiency has improved by 40% each year, according to the Stanford report. Moreover, open-weight models that are free to use are closing the gap with closed models in performance.<\/p>\n But these headline numbers tell only part of the story.<\/p>\n Although the cost of the models has dropped since 2022, the overall cost of ownership \u201chas been resistant to declines,\u201d said Muath Juady<\/a>, founder of SearchQ.AI. \u201cThe real expenses lie in the hidden infrastructure, including data engineering teams, security compliance, constant model monitoring, and integration architects necessary to connect AI with existing systems.\u201d<\/p>\n For every dollar spent on AI models, businesses are spending five to $10 to make the models \u201cproduction-ready and enterprise-compliant,\u201d Juady told PYMNTS. \u201cThe integration challenges tend to be more expensive than the technology itself and require substantial investment in change management and process redesign, which many organizations underestimate.\u201d<\/p>\n Moreover, the cost of AI deployment \u201cis not a one-time expense but an ongoing operational commitment,\u201d Juady added.<\/p>\n So why is AI adoption soaring? Juady said, \u201cbusinesses that are successfully adopting AI are not waiting for costs to drop further; they are identifying specific use cases where even current costs can provide a measurable ROI.\u201d<\/p>\n Read also:<\/strong>\u00a0High Impact, Big Reward: Meet the GenAI-Focused CFO<\/a><\/p>\n Self-Hosting Can Lower Costs<\/strong><\/p>\n For many enterprises, early decisions, such as whether to self-host, use the cloud or use third-party infrastructure, can dictate as much as 40% of AI expenses, said Pavel Bantsevich<\/a>, project manager and solutions advisor at Pynest. Cloud-based hosting may be ideal for prototypes, but costs can spike as workloads scale.<\/p>\n Bantsevich said he worked with a U.S. construction company that\u2019s been in business for a century to develop an AI predictive analytics tool and hosted it in the cloud. Infrastructure costs came to under $200 a month. But once it went live and people started using it, costs soared to around $10,000 a month. Switching to self-hosting using Meta\u2019s open-source Llama model instead of the cloud lowered the cost to about $7,000 a month and has remained under control.<\/p>\n In another case, a European retailer client of Bantsevich\u2019s with more than 50,000 employees wanted to implement a computer vision module for self-checkout machines. But the company didn\u2019t want to use the cloud. It self-hosted instead using a small Llama AI model that performed well. Costs came to less than $10 a month per machine. \u201cIf a cloud solution had been selected, the numbers would have gone sky high,\u201d he said.<\/p>\n Bantsevich believes that costs will continue to decline because datasets are more readily available today and cloud providers also have cut rates to retain customers. \u201cIt is likely we shall see AI costs be similar to electricity bills in the near future,\u201d he predicted.<\/p>\n Meanwhile, Bill Chief Financial Officer Rohini Jain<\/a> advised businesses to take advantage of AI that is already embedded in the platforms they use, such as those for invoicing, payments or forecasting, rather than adding standalone tools with \u201cuncertain\u201d pricing. \u201cIntegrated solutions typically offer better ROI and more predictable costs, such as subscription pricing,\u201d she said.<\/p>\n Fergal Glynn<\/a>, CMO and AI security advocate of Mindgard, said deploying AI can cost as little as $10,000 for basic projects, while large-scale enterprise systems can run into millions of dollars. Most companies spend between $50,000 and $500,000 for practical use cases like analytics tools or chatbots; smaller firms often pay less by using off-the-shelf AI.<\/p>\n Nicole DiNicola<\/a>, global vice president of marketing at Smartcat, told PYMNTS that adopting AI doesn\u2019t have to be \u201call or nothing.\u201d<\/p>\n \u201cMany platforms, including free or low-cost options, make it easy for organizations to start small and scale their adoption over time,\u201d DiNicola said. \u201cUnlike legacy SaaS, which often requires lengthy onboarding, upfront costs, and full-scale deployment to show value, AI can deliver meaningful impact without being fully integrated organization-wide.\u201d<\/p>\n DiNicola pointed to teams embedding AI into workflows and already gaining efficiencies and cost savings. \u201cAI tends to compound in value, but even small-scale adoption can drive clear and measurable improvements.\u201d<\/p>\n A worse outcome would be letting the cost and complexity of AI scare a business into avoiding AI deployment in the first place.<\/p>\n \u201cInaction is often the more expensive path, even if it\u2019s less obvious upfront,\u201d DiNicola added. \u201cWhile that delay might feel safe, early adopters are already building momentum, improving processes, learning faster, and expanding their competitive advantage.\u201d<\/p>\n Read more:<\/strong><\/p>\n How to Choose Between Deploying an AI Chatbot or Agent<\/a><\/p>\n Small Business, Big AI: How SMBs Are Leveling the Playing Field With Enterprise Giants<\/a><\/p>\n![\"graphic](\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/CFO-AI.png\")
<\/p>\n