Does the text format of a spreadsheet change what a model can do with it?
Early results, 9 October 2026. These are not final benchmark numbers.
Setup
14 workbooks, each with a fixed set of questions and ground truth. Each workbook was given to a model as CSV or as Cellprose (the v0.2 Markdown-table text), two repeats, one effort level per model. Seven models from two vendors. Tokens are priced as uncached.
Results, CSV vs Cellprose
Accuracy change is in percentage points. Token changes are relative to CSV.
| Model | Accuracy | Input | Output | Input + output | Cost |
|---|---|---|---|---|---|
| haiku45 | +8.8 | +27% | -27% | -18% | -25% |
| haiku55 | +8.9 | +24% | -31% | -1% | -20% |
| sonnet55 | +0.6 | +24% | -16% | +12% | -3% |
| opus55 | +1.9 | +24% | -18% | +14% | -2% |
| gpt6luna | +3.4 | +26% | -21% | +6% | -11% |
| gpt61sol | +0.5 | +26% | -20% | +17% | +1% |
| gpt6astra | +2.2 | +26% | -16% | +18% | +3% |
Cellprose was at least as accurate as CSV on every model. It uses about a quarter more input tokens and cuts output, mostly reasoning, by 16 to 31 percent. Total tokens fall only for Haiku 4.5 and are level for Haiku 5.5. Cost falls for five of seven models and is level for the two GPT models.
Limitations
- Two repeats, two encodings and one effort level per model. Spread is not yet reported.
- GPT prices are unverified, and token counts are not comparable across vendors.
- This is the earlier v0.2 text. The publishable run follows a stricter protocol (held-out test split, real-world workbooks, confidence intervals).