TL;DR
ScaleDown Summarize is a good, very cheap summarizer for whole documents. On 12 real SEC filings, YouTube transcripts, web pages and Reddit and X posts it was accurate on 11, at a total cost of $0.0061. NexusTrade is moving its document summaries to it.
It is unreliable on numeric tables. Asked to answer stock screener results, it was about 15x faster than Muse Spark 1.3 at about a third of the cost, but it gave a wrong or empty answer on 6 of the 13, so Muse Spark 1.3 keeps that job.
- Usable screener answers, ScaleDown Summarize
- 7 of 13
- Usable screener answers, Muse Spark 1.3
- 13 of 13
- Median response time
- 1.9s vs 28.2s
- Cost per 1,000 screener answers
- $0.44 vs $1.48
How well does ScaleDown Summarize summarize documents?
- Accurate summaries
- 11 of 12
- Median response time
- 0.6s
- Cost for all 12 documents
- $0.0061
- Same summary on a second run
- 11 of 12
Every article, video, post and filing NexusTrade ingests gets a short summary that its AI agent sees before opening the full text. Until now that summary was the document's title plus its first sentence. On scraped pages, X posts, filings and caption transcripts the first sentence is often markup, tag soup or the whole transcript.
ScaleDown Summarize read documents from 1,774 to 182,769 characters, from YouTube, the web, Reddit and X. Its one wrong summary was a short X post saying an option went from 40 cents to $13, which it described as a prediction.
It ran past the 280-character limit on 3 of 12. Trimming the Alphabet summary to 280 characters would have cut its revenue and net income, so NexusTrade asks the model once to shorten any summary that runs long. That took the Alphabet summary from 515 characters to 100 with every figure kept, and anything still too long is trimmed. If a call fails, NexusTrade keeps the first-sentence summary.
Alphabet 10-Q filed April 30, 2026 (182,769 characters)
First-sentence summary: "GOOGL 10-Q filed 2026-04-30: goog-20260331FALSE2026Q10001652044--12-31P1Y0M00Dhttp://fasb.org/us-gaap/2025#Revenues…" ScaleDown Summarize: "Alphabet (GOOGL) Q1 2026 revenue rose 22% to $109.9B; net income surged 81% to $62.6B ($5.11/share)."
A Nancy Pelosi trade tracker page
First-sentence summary: "Nancy Pelosi Stock Tracker: Live Trades & Portfolio 2026 | Congress Stock Tracker: Disclosed trades 322 Est." ScaleDown Summarize: "Nancy Pelosi's household portfolio gained 20.1% in 2025, outperforming the S&P 500, with disclosed trades largely executed by husband Paul Pelosi in tech names like Apple, Nvidia, and Microsoft using long-dated call options."
An X post from @smartertrader
First-sentence summary: "Mstr. 3 layers. Who does this. 110 40 c to 14 120 50c to 5 130 next week 50c to 5. Lolol. Omg 400 contracts each: ## Post [Log in](https://x.com/i/jf/onboarding/web?mode=login&redirect%5Fafter%5Flogin=%2Fsmartertrader%2Fstatus%2F2090756430798197170)…" ScaleDown Summarize: "Sam Parikh notes Mstr. 3 layers: 110 40 c to 14, 120 50c to 5, 130 next week 50c to 5. Omg 400 contracts each."
Where ScaleDown Summarize gets numbers wrong: tables
Show the table · 7 rows, 4 columns
Scroll sideways to compare all columns. Select a heading to sort.
| AVGO | 45.10 | $38.3B | 34.69 |
| GOOGL | 17.61 | $244.2B | 34.69 |
| IONQ | N/A | $-1.4B | 34.69 |
| MSFT | 29.71 | $133.7B | 34.69 |
| NVDA | 28.65 | $192.9B | 34.69 |
| QUBT | N/A | $-15.0M | 34.69 |
| RGTI | N/A | $-238.7M | 34.69 |
No matching rows. Clear the filter to see all records.
Four of the nine columns both models received for a seven-stock technology comparison, as of the October 9, 2026 close. The last column is the same for every row because it describes the whole sector.
After the stock screener runs its query, a second model turns the result table into the answer the user reads. Every number in that answer has to come from the right row and column.
Muse Spark 1.3 ranked profitability correctly: GOOGL at $244.2B of net income, then NVDA at $192.9B, MSFT at $133.7B and AVGO at $38.3B. ScaleDown Summarize said MSFT showed the strongest profitability at $133.7B. It also read the last column, the share of technology stocks above their 50-day average, as each stock's own position, and wrote that all seven sat at 34.69%.
How accurate is ScaleDown Summarize on stock screener tables?
Wrong means the answer stated a value or label the table contradicts, or a fact the table does not contain. Empty means it did not answer.
The 13 answers cover 7 distinct questions, because some were asked more than once. ScaleDown Summarize did well on single-stock snapshots with a handful of fields, such as GRAB, SDGR and TGB. Its errors came on wider tables where several similar numbers sit side by side.
Share classes swapped, twice
On a top-five market cap screen it labeled GOOGL as Alphabet Class C and GOOG as Class A. GOOGL is Class A. The table lists only the tickers.
An invented price move, twice
On a single-stock TGB screen it wrote that the stock had not moved intraday. The data holds only end-of-day prices.
Columns misread, once
This is the seven-stock comparison shown above.
No answer, once
On the same seven-stock question asked again, it returned two sentences restating the question.
ScaleDown Summarize speed and price: about 15x faster, about 3x cheaper
- Median response time, ScaleDown Summarize
- 1.9s
- Median response time, Muse Spark 1.3
- 28.2s
- Cost per 1,000 answers, ScaleDown Summarize
- $0.44
- Cost per 1,000 answers, Muse Spark 1.3
- $1.48
ScaleDown's cost is what OpenRouter billed for these 13 calls. Muse Spark's cost and response time are what NexusTrade recorded for the same 13 production calls. ScaleDown Summarize bills $0.04 per million input tokens and nothing for output.
Can ScaleDown Summarize compact long AI agent conversations?
NexusTrade's AI agent compacts its history when a research run gets long. On 3 real runs with 164,658 to 476,381 characters of history, ScaleDown Summarize wrote coherent summaries in 7.4 to 10.8 seconds that kept the portfolio IDs, numbers and remaining steps. Each one left out a point the production summary kept, such as a strategy that opened and closed the same option in a loop.
Compaction also needs structured fields, such as each portfolio's status and suggested follow-up questions. ScaleDown Summarize accepts only a token limit, with no structured output mode, so the production summarizer stays for this job.
How the test was run
Test method
Freeze real screener requests
Pull the most recent successful Stock Screen Summarizer calls from NexusTrade's request log, with the exact messages the model received and the answer the user got.
13 production calls from October 12 and 13, 2026, all served by meta/muse-spark-1.3-contributor. Calls made by test scripts were excluded.
Send the identical input
Send each call's system prompt and result table to scaledown/summarize on OpenRouter, unchanged.
Both models read the same instructions, table and answer facts.
Grade against the table
Check every number and label in each answer against the table that model was given.
Each answer is usable, wrong or empty.
Summarize ingested documents
Take the 3 most recent ingested documents of at least 1,500 characters from each of YouTube, the web, Reddit and X, and ask for a summary of at most 280 characters. Check each against the full text.
12 summaries, each accurate or wrong.
What this test does not show
A small sample
13 screener answers to 7 questions, logged over two days. A larger sample could move the rates in either direction.
Graded by hand
One reviewer checked each answer against its table. The grades and reasons are kept with the results.
One run per answer
ScaleDown Summarize's wording changes between runs: replaying the same 13 inputs gave an identical answer on 1. The same failures came back on the second run, including the swapped share classes, the invented intraday claim and misread columns.
Volume labels
Some screener tables print share volume with a dollar sign. Muse Spark 1.3 twice called that figure dollar volume. Those were not counted as errors because the table labels them that way.
No comments yet.