Opetek's production AI architecture for capital markets cut measured LLM inference costs by around 90% in a benchmark by separating reasoning from computation, with broader implications for data-intensive AI systems across science, engineering and finance. In an illustrative 30-currency-pair scenario, the same benchmark economics imply a theoretical annual saving approaching $100 million at institutional scale.
The economics of artificial intelligence are changing. For much of the generative AI boom, attention has focused on model capability. Bigger models, larger context windows and increasingly sophisticated reasoning. But as AI moves from experimentation into production, another question is becoming increasingly important. How much computation does all that intelligence consume?
The cost of individual tokens has fallen sharply, but the number of tokens being consumed is increasing as enterprises deploy more agents, longer-running workflows and more sophisticated reasoning systems.
The issue is larger than software pricing. Every unnecessary token represents computation. At sufficient scale, that ultimately means more GPUs, infrastructure, electricity and capital.
The challenge for the next generation of enterprise AI may therefore not simply be building systems that reason over larger data sets. It may be building systems that reason efficiently.
This becomes particularly important when large language models are asked to reason over large numerical datasets.
Consider scientific research. A question in genomics might involve millions of genetic observations. Physics and climate modelling routinely operate across enormous numerical datasets. Engineering systems generate streams of telemetry at frequencies no human could meaningfully inspect. Capital markets has the same characteristic.
A seemingly straightforward question might require years of historical observations across multiple instruments and asset classes, including long histories and high-frequency market data.
If the underlying dataset is brought into an LLM's context, the model has to process those tokens before it can reason about them.
As the dataset becomes larger or more granular, the economics can deteriorate rapidly.
This is not a theoretical concern. The cost of an agentic workflow is determined not just by the user's prompt but by the tools available to the model, the calls it makes and the data returned during the workflow.
At Opetek, we have been investigating a related question at the analytical layer. For a quantitative problem involving a very large dataset, how much of that data needs to enter the reasoning model at all?
ARIUS, Opetek's AI reasoning platform for capital markets, is designed around a core architectural principle, to separate reasoning from computation.
Rather than requiring the model to process the full underlying dataset directly, ARIUS is designed so that the reasoning layer operates on the outputs of quantitative analysis rather than on every underlying market-data point.
The model can then reason over the relevant analytical outputs rather than every underlying observation. The distinction is important.
A human quantitative analyst would not manually read hundreds of thousands of market observations and perform various arithmetic and other operations.
They would understand the question, design the analysis, use computational tools to process the data and then interpret the result.
AI systems can work the same way. ARIUS is designed to reason across market data, pricing models and other financial information while producing reproducible and auditable quantitative outputs. This is part of Opetek's broader approach to building reasoning systems specifically for capital markets.
To understand the economics, Opetek ran a benchmark using 10 years of EUR/USD historical data to calculate realised volatility.
In the first approach, the underlying time series was passed directly into the LLM context.
In the second, the same analytical task was performed using Opetek's production architecture, designed to keep bulk market data outside the reasoning layer.
The difference was substantial.
That represents a saving of approximately $0.732 per query, around a 90% reduction in measured LLM inference cost for the benchmark.
This result is specific to the benchmark described above and should not be interpreted as a claim that every AI workload will achieve the same cost reduction. It demonstrates a broader architectural point that where computation happens can materially change the economics of reasoning systems.
A single query costing $0.80 rather than $0.07 is economically irrelevant. The picture changes when AI becomes part of the daily operating environment of a large institution. Take an illustrative front-office user making 40 comparable analytical requests per day.
Based on the benchmark:
| Direct-in-context approach | Opetek architecture | |
|---|---|---|
| Cost per query | $0.802 | $0.070 |
| Cost per user/day | $32.08 | $2.80 |
| Cost per user/year* | $7,218 | $630 |
*Assuming 225 working/trading days.
That implies approximately $6,588 of annual LLM cost reduction per user under those assumptions.
Now consider an illustrative deployment across 500 front-office users spanning trading, sales, structuring and quantitative functions.
Annual model cost would be approximately:
The difference is approximately $3.3 million per year.
These are scenario calculations, not observed customer savings. But the assumptions are explicit, making the economics straightforward to inspect.
The EUR/USD test is deliberately simple. Institutional capital-markets questions can be much more demanding.
A trader might ask for analysis across dozens of currency pairs rather than one. A volatility question might require multiple tenors and strikes across complete volatility surfaces. Cross-asset analysis might combine rates, FX, equities and credit. Intraday questions can replace daily observations with far more granular data.
More sophisticated requests can also require deeper reasoning and multiple analytical stages. In architectures where the underlying dataset is repeatedly included in the model context, each additional reasoning step can require the model to process that data again.
The economics can therefore compound in two dimensions. More data per analytical request, and more model interactions as the system reasons through the problem. By keeping bulk numerical data outside the reasoning layer, the Opetek architecture is designed to avoid repeatedly passing the same underlying dataset through the LLM during multi-step reasoning.
With the Opetek architecture, the bulk numerical workload remains outside the LLM.
This means the economic difference can widen as workflows require larger datasets, more granular data and deeper reasoning.
To illustrate the potential scale, consider a trader comparing realised volatility across 30 currency pairs rather than analysing EUR/USD alone.
A direct-in-context architecture would need to expose the reasoning model to substantially more underlying time-series data. If, for illustration, we assume that model-facing inference cost scales approximately linearly with the amount of equivalent historical data being passed into the model, applying the benchmark economics across 30 currency pairs produces indicative annual economics of:
These are illustrative scenario calculations, not observed customer savings. The calculation assumes 500 users, 40 comparable analytical requests per day and 225 trading days, and applies the measured EUR/USD benchmark economics across 30 equivalent datasets.
It also assumes approximately linear scaling for the purpose of illustration. Actual model costs would depend on the model, workflow, dataset, context construction and the amount of analytical output returned to the reasoning layer.
The significance is not the precise $98.8 million figure. It is that once AI is asked to analyse many instruments, large historical datasets and repeated institutional workflows, architectural choices that appear economically minor at the level of a single query can compound dramatically.
Much of the AI industry's response to inference economics has focused on making models themselves cheaper.
That work matters. Model providers are competing aggressively on inference pricing and token efficiency, while researchers are developing architectures specifically intended to reduce unnecessary token consumption.
But model efficiency is only one layer of the problem. There are at least three places where AI systems can become more efficient.
The third question may become increasingly important as AI expands beyond language-centric tasks into scientific, engineering and quantitative domains.
The architectural mistake may be treating a reasoning model as though it were also the optimal place to perform every piece of computation.
LLMs are extraordinarily capable at understanding ambiguous questions, decomposing problems, choosing analytical approaches and interpreting results.
Conventional computers are extraordinarily capable at performing numerical operations across millions of observations. Those are different capabilities. Using expensive model inference to perform work that conventional computation can handle efficiently is unlikely to be the optimal architecture at scale.
The objective should therefore not be to minimise reasoning. It should be to spend reasoning where reasoning creates value.
That principle is central to Opetek's approach. Trusted AI reasoning systems that allow financial institutions to make better, faster and more defensible decisions by embedding advanced reasoning into quantitative workflows.
Capital markets is an unusually demanding environment for AI. Datasets are enormous. Questions are quantitative. Latency matters. Accuracy matters. And the financial incentive to analyse more information, more frequently, is substantial.
That makes financial markets an effective proving ground for architectures that combine reasoning models with conventional computation.
But there is nothing uniquely financial about the underlying principle. The same issue appears wherever sophisticated reasoning meets enormous datasets, biology, physics, medicine, climate science, engineering and potentially almost every scientific field adopting foundation models.
The broader question is therefore not whether AI will reason over increasingly large amounts of data. It almost certainly will. The question is whether every piece of that data needs to become a token before useful reasoning can take place.
Our results suggest the answer is no. And if AI adoption continues at its current pace, that distinction could become worth far more than the cost of a few tokens.
