GPT-6 Luna on Lucy data matches 98% of GPT-6 Astra’s quality for 0.22% of the cost
Same filings, same questions, three models. The only change is what the model reads: Lucy data or the full filing.
Only the passages that matter, picked from a filing Lucy has split into Items
The whole filing, converted to text with an open-source tool (markitdown)
Extraction with one right answer
- Headcount: total, as-of date, and any breakdown by region or function
- Major customers (10%+ of revenue): name, share, revenue or receivables basis, period
A good extraction finds every value, skips look-alikes (a subsidiary's headcount, say), and gets each value right.
A full filing averages about 190,000 tokens, yet the headcount usually sits in one or two paragraphs. With full filings, over 95% of the cost is the filing text going in.
What goes into Lucy data
- Split by ItemEach topic reads only the Items where its answer can be.
- Tables kept intactEvery number stays attached to its row, column, and unit.
- Search inside ItemsTens of thousands of tokens, not hundreds. Even the largest 20-F fits in one call.
- Rules per topicWhat counts as company-wide headcount, when a customer share is in scope.
| Filings | 60 US annual reports, fiscal 2025, picked at random (40 10-K, 20 20-F) |
| Conditions | 2 inputs × 3 models = 6, same instructions for all. Luna + Lucy data also had a few check lines for commonly missed items. |
| Long filings | Full filings over the 1M-character input limit were split into 2–3 chunks and merged |
| Reasoning | Luna high · Sol and Astra default (medium) |
| Cost | Actual tokens × list price, no cache discount |
| Grading | Blind AI grader (Opus); every item marked wrong rechecked by a second grader |
01Cost: Lucy data cuts every model's cost to about 1/5
02Quality: Lucy data holds it, and lifts Luna
- Sol, Astra: Lucy data and full filing within 1 point, inside run-to-run swing.
- Luna: +6 points with Lucy data.
- Core values (headcount, date, customer, share) were 99.7–100% correct in every condition; the few wrong values (5 in total) came mostly from full filings.
03Big full filings didn't fit in one call
9 of 60 full filings went over the input limit, including 8 of the 20 20-Fs. They needed extra code to split, call several times, and merge. Every Lucy input fit in one call.
- 98% · 1/459With Lucy data, low-cost Luna gets 98% of frontier Astra's quality at 1/459 the cost.
- ≈ 1/5Sol and Astra keep their quality on Lucy data at about 1/5 the cost.
- 9 of 60full filings were too long for one call and needed split-and-merge code. Every Lucy input fit.
Read as text
GPT-6 Luna on Lucy data matches 98% of GPT-6 Astra’s quality for 0.22% of the cost
Same filings, same questions, three models. The only change is what the model reads: Lucy data or the full filing.
98% Luna + Lucy data quality, relative to Astra + Full filing 1/459 Its cost, relative to Astra + Full filing ≈ 1/5 Cost on Lucy data for the same model, quality held November 9, 2026 · 60 filings · 2 topics · 6 conditions Two inputs Lucy dataOnly the passages that matter, picked from a filing Lucy has split into Items
Full filingThe whole filing, converted to text with an open-source tool (markitdown)
Three models · input price per token Luna Low-cost gpt-6-luna 1× Sol Premium gpt-6.1-sol 20× Astra Frontier gpt-6-astra 100×Extraction with one right answer
- Headcount: total, as-of date, and any breakdown by region or function
- Major customers (10%+ of revenue): name, share, revenue or receivables basis, period
A good extraction finds every value, skips look-alikes (a subsidiary's headcount, say), and gets each value right.
A full filing averages about 190,000 tokens, yet the headcount usually sits in one or two paragraphs. With full filings, over 95% of the cost is the filing text going in.
The inputWhat goes into Lucy data
- Split by Item Each topic reads only the Items where its answer can be.
- Tables kept intact Every number stays attached to its row, column, and unit.
- Search inside Items Tens of thousands of tokens, not hundreds. Even the largest 20-F fits in one call.
- Rules per topic What counts as company-wide headcount, when a customer share is in scope.
| Filings | 60 US annual reports, fiscal 2025, picked at random (40 10-K, 20 20-F) |
| Conditions | 2 inputs × 3 models = 6, same instructions for all. Luna + Lucy data also had a few check lines for commonly missed items. |
| Long filings | Full filings over the 1M-character input limit were split into 2–3 chunks and merged |
| Reasoning | Luna high · Sol and Astra default (medium) |
| Cost | Actual tokens × list price, no cache discount |
| Grading | Blind AI grader (Opus); every item marked wrong rechecked by a second grader |
01 Cost: Lucy data cuts every model's cost to about 1/5
02 Quality: Lucy data holds it, and lifts Luna
- Sol, Astra: Lucy data and full filing within 1 point, inside run-to-run swing.
- Luna: +6 points with Lucy data.
- Core values (headcount, date, customer, share) were 99.7–100% correct in every condition; the few wrong values (5 in total) came mostly from full filings.
03 Big full filings didn't fit in one call
9 of 60 full filings went over the input limit, including 8 of the 20 20-Fs. They needed extra code to split, call several times, and merge. Every Lucy input fit in one call.
Takeaways- 98% · 1/459With Lucy data, low-cost Luna gets 98% of frontier Astra's quality at 1/459 the cost.
- ≈ 1/5Sol and Astra keep their quality on Lucy data at about 1/5 the cost.
- 9 of 60full filings were too long for one call and needed split-and-merge code. Every Lucy input fit.