"Small" is doing too much work in this conversation. It is read as a size class, a kind of model, something you can put on a slide and buy. The evidence people cite for it says something different. Gartner's forecast is about "small, task-specific" models, and the vendor results we are about to look at are about task-specific models too. The word that carries the saving is the second one.
A small model that has not been shaped to a job is just a weaker model. What pays is a model fitted to a narrow, repetitive task and run at a volume where the price per call matters. Size is how that fit usually gets cheap, not why it works.
The label has been fading anyway. In US Google Ads data pulled through DataForSEO in October 2026, small language models ran at 4,400 searches a month in September 2025 and 2,400 by July and August 2026, a drop of about 45%. It still averages roughly 2,900 a month, at $1.64 a click. We think the label should go while the idea stays, and the idea never depended on the word.
What the numbers say, and who produced them
Gartner predicts that by 2027 organizations will use small, task-specific models at volumes at least three times those of general-purpose LLMs. It also projects domain-specific generative AI models growing from about 1% of enterprise use in 2024 to more than half by 2027. Forecasts of this kind age badly, so hold the numbers loosely.
The benchmark evidence is stronger on direction than on size. ScaleDown, an AI infrastructure firm, reported through Forbes that its task-specific small models ran about 8% more accurate, 161 times cheaper and 3.8 times faster than Anthropic's Claude on classification and summarization jobs. Against OpenAI's models the figures were roughly 8.7% more accurate, 89 times cheaper and 2.4 times faster. Against Gemini: about 9%, 29 times and 8.3 times. In one example, 10,000 summaries a day cost about $7.20 on a task-specific model and $58 on GPT-4.1 Mini, with human evaluators unable to tell the outputs apart.
Read who produced it. ScaleDown sells small models. The test covers classification and summarization only, and the baselines are general models doing narrow work, which is the fight a task-specific model was built to win. We would not quote 161x to a finance director. We would quote the shape: on bounded tasks, a model tuned to the job can match a frontier model at a small fraction of the price. At the example's own figures the gap is $50.80 a day, about $18,500 a year, for one workflow. A real number, though not a dramatic one for a single process. It starts to matter when you have thirty processes.
So what to buy
Start with the workload, not the model. If you cannot say what the task is, how often it runs and how you will know it was done right, no model size will help. The saving also needs volume. At a few thousand calls a month the choice of model is a rounding error, and the engineering time to run a second model costs more than it saves. At millions of calls it becomes a budget line, and the work that reaches that volume is back-office work: classifying tickets and invoices before they are routed, pulling fields from documents into your ERP, summarizing records and delivery reports, flagging unusual transactions. None of it needs a model that can also write a sonnet.
So separate the narrow, repetitive tasks from the judgment work, price both paths at your real volume with finance in the room, and test on your own documents. A vendor's benchmark is a hypothesis about your data, not a finding. The general model keeps the open-ended requests and the tuned one takes the repeatable steps underneath, a division of labor and not a replacement.
ThinqHub implements Odoo and ERPNext, and the high-volume document and record work above lives inside those systems, which is why we ask about the workload first. Gartner's forecast that 40% of agentic AI projects will be canceled by 2027 is covered in our piece on agentic AI implementation. The same fit-the-tool-to-the-task discipline runs through our approach, and you can talk to us if you want a workload audited.
