Report Generation with DAI Studio
Date: September 7, 2026
The point of this one isn’t the model — it’s the context. I built a single workflow diagram once, then ran it through five different models to see what “same input, different model” actually costs and produces.
The diagram — SP500 PE Monthly — is three tasks: SP500 PE monthly (fetch the data) → Chart the Data → Build Report, with a red Rule: HTML annotation pinning the output format. That’s it. No hand-written spec, no back-and-forth.
The entire human input was two prompts against the diagram’s raw XML:
Plan mode:
plan out building https://dai.studio/blog/sp500-pe-report/raw
Code mode:
build
Five models, five reports, five prices:
| Model | Cost | Report |
|---|---|---|
| Claude Opus 5 | $2.65 | View Report — compact hand-built SVG chart plus a full stats table (percentiles, mean, max/min with dates, current percentile rank) |
| GPT-5.6 Sol | $1.90 | View Report — the most editorial output: headline treatment, a written interpretation paragraph, and an explicit methodology note (R-7 interpolation) |
| GLM 5.3 | $0.86 | View Report — concise, with a real analytical findings paragraph (decade-by-decade averages, the dot-com era’s 80-month streak above the high band) and a correct citation back to the source workflow |
| Kimi K3 | $0.42 | View Report — correct numbers, but the chart is brute-forced: every gridline and data point hardcoded inline, ballooning the file to over 1,300 lines for the same chart the others drew in a fifth of the code |
| Claude Sonnet 5 | $0.31 | View Report — the smallest correct output, clean custom SVG, same numbers as the expensive runs |
All five models agree on the numbers that matter: 1,281 monthly observations from January 1920 to September 2026, an 80th-percentile PE of 22.12, a median of 16.52, a 20th percentile of 10.42, and a current reading around 26.36 — roughly the 90th percentile of the entire series.
Cost didn’t track quality here so much as verbosity. Claude Sonnet 5 at $0.31 matched the expensive runs on substance while using a fraction of their tokens; Kimi K3 spent effort hardcoding SVG coordinates for a chart the others drew far more efficiently. Same diagram, same two prompts, five different ways of getting to the same correct answer.
Cheers,
Ted :)