You've probably copied an entire sales spreadsheet into ChatGPT, only to end up with a prompt that ballooned in length and an answer that stayed frustratingly vague. Feed it hundreds of raw numbers and you'd expect the AI to spot the trend on its own — instead you get a wishy-washy summary and a token bill that's shot up.

The catch is that numbers are an unusually expensive input for language models. Models like ChatGPT break input into chunks called tokens, and a string of decimal-heavy figures inflates that token count fast. A new arXiv paper tackles this head-on and lands on a surprisingly simple fix: don't hand the model numbers as text — hand it a chart image instead.

The Study

The researchers start from telecom network analysis. 4G and 5G base stations spit out multiple key performance indicators (KPIs) every 15 minutes, and when that stream of numbers is fed into a language model as text, it balloons into thousands of floating-point tokens. According to the paper, LLM inference already accounts for more than 90% of AI operating energy, and that energy draw scales with the number of input tokens. So the researchers asked a pointed question: what happens if you take the same time-series numbers, turn them into a 2D plot — a chart image — and hand that to a vision-language model (VLM) instead?

What They Found

The results were striking. Across three models — Llama-3.2-90B, Qwen2.5-VL-72B, and Pixtral-12B — input tokens dropped 3.6 to 10.4 times, and measured inference energy fell 1.8 to 2.5 times. In a telecom edge environment monitoring 200 base stations every 15 minutes, that works out to roughly 7.2 megajoules of energy saved per day.

The more interesting finding is accuracy. A fine-tuned Llama-3.2-90B-Vision achieved 220.7% higher precision than the same model working from text alone, and beat traditional anomaly-detection methods like LSTM and ARIMA by more than 144% on telecom anomaly detection. On public benchmarks, Pixtral-12B held an average F1 of 0.82 while improving its energy-to-performance ratio by 20.6 times. Once the number of tracked indicators reaches 24, the text representation blows past the 128K-token limit of most commercial models — making the task literally impossible without truncation. The image representation, by contrast, stays comfortably within standard limits. Cheaper and more accurate at once is a rare combination.

Putting It to Work

Try changing just one habit the next time you analyze Excel data with ChatGPT. Instead of pasting in raw numbers — monthly revenue, ad performance, visitor trends — build a chart in Excel, capture it as an image, and upload that instead. Keep the axis labels and date range visible; don't crop them out. Every visual cue you leave in is one more clue the model can read directly off the chart. The telecom KPIs in the paper and a solo entrepreneur's sales table differ only in scale — both are, fundamentally, numeric time series with multiple metrics moving over time.

The payoff comes in two forms. First, fewer tokens means a direct cost cut if you're paying by the API call, and if you're chatting interactively, it means a long table is less likely to crowd out your conversation context and make the model forget earlier instructions. That matters most in analysis work that spans months of data across an extended conversation. Second, when you're asking about trends, spikes, or anomalies, there's now evidence that a chart draws a better answer out of the model than a table does. The real trick to ChatGPT Excel analysis isn't reciting numbers — it's asking the model to read a picture. Try something like: "Point out the odd stretch in this chart and suggest a hypothesis for what caused it."

The Fine Print

It's worth being honest about the limits here. This study covers anomaly detection and analysis of numeric time series like telecom KPIs. For tasks that need the raw numbers themselves — exact sums, looking up a specific value — a table is still the right tool. And that 220.7% precision gain was measured on a specific fine-tuned model, so you shouldn't expect the same magnitude of improvement from an off-the-shelf chatbot. Still, the direction is clear: if what you want is the trend, one picture is cheaper than a thousand rows of numbers.