Purpose-built AI model for spreadsheet editing.
DAX-1 is a specialized AI model for precise, executable spreadsheet edits with lower latency and serving cost.
Dax-1 goal
Dax-1 is Decide's deterministic spreadsheet-editing model, built by post-training a Qwen3-14B base model for exact workbook repair. It is designed to handle common Excel tasks like fixing formulas, completing lookups, cleaning duplicates, correcting rows, and repairing sorted or filtered tables.
Dax-1 produces executable spreadsheet edits instead of free-form prose, making it faster, easier to verify, and cheaper to run. On our frozen 350-task benchmark, Dax-1 reaches 91.7% strict workbook accuracy, with 1.48s p50 latency and about 90.8% lower benchmark cost than Opus 4.5 under a $1/hour warm-GPU serving assumption.
On this page, Dax-1 refers to the production configuration used in Decide: a quantized spreadsheet-editing model with deterministic patch execution and workbook-state verification.
Frontier accuracy at a fraction of the cost.
Price-performance
Frontier accuracy at a fraction of the cost.
Dax-1 reaches 91.7% accuracy at a $0.69 benchmark run cost. The cost assumes the quantized production system on a warm GPU at $1.00/hour.
Latency
Fast enough for repeated spreadsheet work.
Measured latency is 1.48s at p50 and 2.42s at p95, keeping everyday workbook edits responsive without giving up deterministic execution.
Accuracy
91.7% strict accuracy on deterministic spreadsheet editing.
Dax-1 passes 321 of 350 strict workbook tasks, slightly ahead of GPT-5.5 and well ahead of Fable 5 on this frozen evaluation.
Try Dax-1 in Decide
Put Dax-1 to work in Decide.
Dax-1 powers the Decide Agent across the web app, Excel, Google Sheets, and API. Put it to work on your next workbook.