- Opetek's CEO argues AI models should not read every data point. In his firm's test, moving one calculation outside the model cut a query's cost by about 90%.
- It is a vendor's own benchmark on one task, and it reports cost only. The author says it does not mean every AI workload gets 90% cheaper.
- Ask your team which AI workflows feed raw data into the model, and which of those tasks are fixed calculations that ordinary software could run.
A bigger context window is a capability. It is not an instruction to fill it. That is the argument in a recent InfoWorld contributed article by Varqa Abyaneh, CEO of Opetek, which builds AI reasoning systems for capital markets. He sells a product in this area, so read his numbers as a vendor's evidence. The idea behind them still deserves attention from anyone who pays an AI bill.
What the article says
A context window is the amount of text a model can hold in view during one request, counted in tokens. A token is a small chunk of text, often part of a word. Abyaneh sketches how quickly these windows have grown. By his timeline, Claude's window went from 9,000 tokens in May 2023 to 200,000 by November 2023. Google's Gemini 1.5 Pro offered two million by June 2024.
He adds that a million tokens had reached general availability across Anthropic's Claude models by March 2026. At least thirteen hosted frontier models now match or exceed that size, he says.
Cost is the catch. Abyaneh cites Ramp data showing that what US businesses pay per million tokens dropped 41% between its March 2026 peak and September 2026. His point is that cheaper tokens do not settle the matter. In data-intensive fields, he argues, the extra volume of data pushes spending up faster than the discount can offset.
The test behind the claim
Opetek ran a benchmark on its own proprietary reasoning architecture. It asked for realized volatility, a standard measure of how much a price has swung, across a decade of euro-to-dollar exchange rates.
When the raw time series went into the model's context, Opetek says each query cost roughly $0.80. When the calculation ran outside the reasoning layer, the cost was about $0.07. That is a reduction of approximately 90%.
Treat the result with care. It is the vendor's own test. It covers one task. The article gives cost figures only, with no comparison of answer quality. The author is explicit that the result should not be stretched to all AI workloads.
How the split works
The design separates two jobs. The model works out which calculation a question needs. A conventional program runs it. The model then interprets the figures that come back.
This works because many calculations are deterministic. The same input always gives the same output, so ordinary software can run them cheaply. When the sum is fixed and cheap to run elsewhere, Abyaneh sees little value in routing every data point through the pricier model.
His analogy is a human quantitative analyst. Such an analyst does not work through hundreds of thousands of observations by hand. The analyst picks the calculation, runs it in a tool, and thinks about the output.
He divides AI efficiency into three layers. The first is how cheaply the model itself produces answers. The second is how well the system fetches the information the model needs. The third is the design choice about what reaches the model at all.
The argument: input volume can drive the bill
Token price is one half of the sum. The other half is how many tokens you send. For deterministic, data-heavy calculations, this test suggests input volume can dominate the bill. The sources do not show it for other kinds of work.
The cost bites hardest where the same analysis runs again and again, he argues. Abyaneh notes that financial markets pair huge datasets with repeated, highly numerical questions. He says the same challenge applies in genomics, physics, engineering, climate modelling and medicine. WebPulse has not tested those claims.
A large context window is still useful. It lets a model take in a long contract or a messy file. But the ability to read everything is not a reason to ask for it on every query.
Questions for your team
Ask which AI workflows pass raw tables, logs or time series straight into a model. Then ask which of those tasks are fixed calculations, such as totals, averages or risk measures, that existing software could run.
Ask for cost per query broken down by what goes in and what comes out. Ask any vendor quoting savings to test on your data and to report answer quality next to price. And decide who owns the choice of what reaches the model, because that choice shapes the bill.
Before you pay a model to think, check what you are paying it to read.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: InfoWorld.
