TL;DR
ScaleDown Extract is fast and cheap, but it is not accurate enough to pull EPS out of earnings releases. It returned the right GAAP EPS on 86 of 138 real filings, against 121 for DeepSeek V4 Flash, while running about 12x faster at about a tenth of the cost.
It did better on fields that appear once in a document, such as the quarter-end date. For a number picked from a dense earnings table, a general model was more reliable in this test.
- Correct GAAP EPS, ScaleDown Extract
- 86 of 138 (62.3%)
- Correct GAAP EPS, DeepSeek V4 Flash
- 121 of 138 (87.7%)
- Median response time
- 2.1s vs 25.6s
- Cost per 1,000 filings
- $0.55 vs $5.40
Where ScaleDown Extract goes wrong: picking one number from a table
The job is narrow. Read an 8-K earnings release the day it is filed and return one number: GAAP diluted EPS for the quarter just reported. A release prints that number beside last year's quarter, the six-month total and the non-GAAP figure, so the model has to pick the right cell.
What ScaleDown Extract was compared against
NexusTrade pulls EPS from every new earnings 8-K with DeepSeek V4 Flash, so backtests and screens see the quarter's EPS the day it is announced, before the 10-Q exists.
That job is already cheap: 19,037 calls since September 13, 2026 cost $40.08 in total. ScaleDown Extract is a purpose-built extraction model billed at $0.04 per million input tokens, so the test asked whether it could match DeepSeek's accuracy for less.
- Production calls since 2026-09-13
- 19,037
- Recorded cost of those calls
- $40.08
How accurate is ScaleDown Extract on SEC filings?
Correct means within half a cent of the EPS in the company's later 10-Q or 10-K.
ScaleDown Extract speed and price: 12x faster and about 10x cheaper
ScaleDown's cost is what OpenRouter billed for these 138 calls, $0.0764. DeepSeek's cost applies production's recorded rate, $0.343 per million prompt tokens with output included, to the 2,172,853 prompt tokens it read in this test.
Does ScaleDown Extract's confidence score help?
ScaleDown Extract returns a confidence score with every value. Raising the cutoff removes most of its wrong answers, and it removes correct ones at a similar rate.
What ScaleDown Extract is good for
Dates were close. ScaleDown Extract matched the quarter-end date on 123 of 138 filings, against 128 for DeepSeek V4 Flash. Its EPS misses came from dense tables where several numbers sit side by side. On eBay it returned the tax rate.
That points to fields that appear once in a document, such as a date, a name or an identifier, where its speed and price are an advantage. For picking one cell out of a dense earnings table, DeepSeek V4 Flash stays in NexusTrade's pipeline.
How the test was run
Test method
Pick filings with a known answer
Sample 600 earnings 8-Ks filed in 2026 whose EPS was later confirmed by the matched 10-Q or 10-K statement.
138 had the original documents archived as text and were scored. They were accepted between July 15 and August 27, 2026.
Rebuild the exact input
Load each filing from the same archive the production worker reads and assemble its text the same way.
Both models read the same press release text.
Ask both models
DeepSeek V4 Flash runs the production SEC 8-K EPS Extraction prompt. ScaleDown Extract gets plain-English descriptions of the same fields: GAAP diluted EPS, GAAP basic EPS, adjusted EPS and quarter-end date.
Every DeepSeek response was served by deepseek/deepseek-v4-flash and every ScaleDown response by scaledown/extract.
Score against the 10-Q
Compare each answer with the statement EPS, allowing half a cent of rounding.
Each filing is correct, wrong or no answer for each model.
What this test does not show
One task
This measures GAAP quarterly EPS from earnings releases. ScaleDown Extract may compare differently on other fields or documents.
Release versus 10-Q
On three filings both models returned the same value and the 10-Q differed, which suggests the release and the later statement disagree rather than a model error.
Recent filings only
Only July and August 2026 releases had archived text, and scanned PDFs were skipped.
No comments yet.