Gemini for SEC filing research
Blind round 2026-11-10 · ranked #6 of 6 · Octo ranked #1 · plans checked Nov 9, 2026
Gemini answered all 40 questions in our blind test and scored 125 of 160, ranked #6 of 6; Signal8's Octo ranked first with 160. Gemini lost the most marks: 4 answers had a wrong figure (1 of them stale), 3 gave no figure, and several cited only a site name such as Stock Titan. 25 of its 40 answers still scored 4 of 4, so its problem is less reading filings than finding and citing the right one. In the harder round 3 its free tier blocked 12 of 40 questions; it scored 80 of 112 on the 28 it answered, with 5 wrong figures, and is not ranked.
What it is, and what it costs
Gemini is Google's assistant. Google's US plan page (gemini.google/us/subscriptions, read Nov 9, 2026) lists Free at $0 with a Google Account (the "Flash-Lite model", Deep Research, Gemini Live and Canvas), Google AI Plus at $4.99 a month with "2x higher usage" than Free, and Google AI Pro at $19.99 a month with "4x higher usage" and access to "Gemini 4 Argon". Both our runs used the default model in the same signed-in browser, which was on the free tier in the pilot.
What we measured in the blind round
Figure correct 33 of 40, cites a source 30, source contains it 26, as-of date right 36. Every answer that lost a mark:
| Question | Company and topic | Marks of 4 | Blind grader's note |
|---|---|---|---|
| R2-07 | WKSP, ATM or warrants | 2 | [STALE] Gives $4.0 million (the original ATM capacity, not the $493,000 in the Aug 2026 supplement) with H.C. Wainwright; names Form 424B5 filed Aug 7, 2026; states Aug 7, 2026. |
| R2-08 | OGEN, ATM or warrants | 0 | [DECLINED] Says it has no access to the 10-Q and gives no figure; only points to EDGAR generally. |
| R2-12 | PULM, ATM or warrants | 0 | [DECLINED] Said it had no access to the filing; only cited the EDGAR homepage. |
| R2-14 | CDT, Reverse split | 2 | Gave 1-for-25 eff. Sept 28 and 1-for-20 eff. Nov 5, 2026 but no URL, accession, or form type named (only filing dates in a table). |
| R2-16 | GOSS, Reverse split | 2 | Gives 1-for-80, effective Sept 10, 2026 at 11:59 p.m. ET. Its only source is 'Stock Titan', with no URL or filing. |
| R2-20 | CYCU, Reverse split | 2 | Gives 1-for-8, effective Aug 28, 2026. Its only source is 'Stock Titan', with no URL or filing. |
| R2-21 | SER, Cash and going concern | 2 | Gave $23.0M at June 30, 2026, but cites only 'Q2 2026 financial reporting filed with the SEC', with no form type, filing date, URL or accession. |
| R2-22 | JSPR, Cash and going concern | 2 | Gave $7.3M but stated it as 'specifically $7,300,000', an exact figure that does not match the key's $7,314,000; it cites the Form 10-Q for the period ended June 30, 2026. |
| R2-24 | CLRO, Cash and going concern | 2 | Gave $75,000 at June 30, 2026, but attributes it to 'StockAnalysis.com' and 'its Form 10-Q filed with the SEC', with no URL, accession or filing date. |
| R2-28 | ISPC, Cash and going concern | 0 | [DECLINED] Said it had no access to the 10-Q data and gave no figure; cited only the EDGAR company search homepage. |
| R2-30 | ADGM, Shares and float | 3 | Gave 22,210,459 but as of June 30, 2026 rather than the cover-page date August 10, 2026; names the Form 10-Q for quarter ended June 30, 2026 (key filing). |
| R2-33 | BTLN, Shares and float | 2 | Gave $0 (invented 'no public trading market' quote) vs key ~$16,769,000; cites the 10-K FY June 30, 2026 filed Sept 28; date Dec 31, 2025 correct. |
| R2-34 | POCI, Shares and float | 2 | Gave ~$25,603,000 as of Dec 31, 2025 but wrongly said the cover gives no non-affiliate share count (key: 6,110,561), so multi-part figure fails; names the 10-K filed Sept 28, 2026. |
| R2-39 | LTRN, Insider purchase (Form 4) | 2 | Keyser and Schalop at 48,544 each, 97,088 total, May 12, 2026; sources are only bare labels (Fintel, FORM 4, GuruFocus) with no URL, accession or filing date. |
| R2-40 | NSPR, Insider purchase (Form 4) | 2 | 21,000 shares at $1.20 on May 12; sources are only bare labels (GuruFocus, MarketBeat) with no URL or filing. |
Blind round caveats: every question named its filing, so the top four tools finished within 2 marks; the graders were Claude instances, the same model family as one contestant; plans differed by tool. Details on the blind round page.
Blind round 3 (2026-11-11): harder questions
Round 3's 40 questions did not name the filing, so each tool had to find the right and latest one. Gemini scored 80 of 112 and was not ranked, because it answered 28 of 40 (its free tier blocked 12 questions); Signal8's Octo ranked first with 156 and no wrong figure. Figure correct 23 of 28, cites a source 17, source contains it 13, as-of date right 27; wrong figures 5. Where its answers came from is not measured, because its links could not be captured from its page. Its round 3 answers were graded blind in a second pass, beside only one other tool's answers.
| Question | Company and topic | Marks of 4 | Blind grader's note |
|---|---|---|---|
| R3-01 | LIDR, ATM, warrants or shelf | 2 | Gave $50.0M under the Sept 15, 2026 At Market Issuance Sales Agreement (correct, prior ATM terminated); only names 'Stock Titan'/'TradingView' with no URL, accession or filing form+date. |
| R3-02 | XXII, Reverse splits | 2 | Gave 5 splits with correct ratios/dates and 1-for-14,904,000 (correct, 1-for-58 excluded); only names 'Stock Titan' with no URL or filing. |
| R3-05 | RNXT, Nasdaq status (8-K) | 2 | Said regained compliance but dated it 'officially ... on August 21, 2026' and hedged the 8-K as 'August 20-21'; key date is Aug 20, 2026, so the stated date is wrong/hedged; cites Form 8-K Aug 20-21 plus 'TradingView'. |
| R3-06 | CODX, ATM, warrants or shelf | 1 | Total 5,012,756 correct but says 3,294,894 May warrants still outstanding (key 1,592,532); cites only 'Stock Titan' with no document; as-of Sept 30, 2026 stated. |
| R3-07 | JAGX, Reverse splits | 2 | Four splits and 787,500 match the key; sources named only as 'Jaguar Health, Inc.' and 'Trendlyne.com' with no URL, accession or dated filing; effective dates stated. |
| R3-08 | SEGG, Cash and runway | 2 | $785,905 at March 31, 2026 matches and correctly excludes the Nov 2 Q2 10-Q; cites only 'Stock Titan' with no document link; balance-sheet date stated. |
| R3-10 | BFRG, Nasdaq status (8-K) | 2 | February 8, 2027 matches; sources named only as 'Stock Titan' and 'TradingView' with no document link; as-of Oct 31, 2026 stated. |
| R3-12 | IVF, Reverse splits | 2 | 4 splits, 1-for-1,440, correctly leaves out the Oct 30 split; names only SEC.gov/Nasdaq Trader/Stock Titan with no specific document, accession or filing date; window stated. |
| R3-16 | VTGN, ATM, warrants or shelf | 2 | Total $175.0M correct but remaining $68.0M is wrong (key $139.5M); cites 10-Q for quarter ended June 30, 2026; source mark 0 because figure 0. |
| R3-17 | NVVE, Reverse splits | 2 | 114,991, 1-for-18, July 6, 2026 all correct; only source is 'Nasdaq Trader' with no URL, form or date. |
| R3-21 | DCOY, ATM, warrants or shelf | 2 | Gives $3.25 (reduced from $5.91), and says Series B's 1,184,434 shares were exercised at $3.25 with 2,368,868 new warrants issued, all correct. Its only attribution is a bare 'SEC.gov' with no URL, accession or form type, so it gets no citation mark. As of September 30, 2026 is stated. |
| R3-22 | KITT, Reverse splits | 2 | Gives 600,233 (or 600,234 rounded) with a factor of 48 from the 1-for-8 (April 21) and 1-for-6 (September 24/25) splits, which is correct. It has no links, and the only source it names is the FY2025 10-K filed April 15, 2026, copied from the question; nothing is cited for either split. The split dates are stated. |
| R3-23 | TELO, Cash and runway | 2 | Cash of $4,515,704 is correct, but the runway is wrong: it says cash is not sufficient for 12 months (through Q3 2027), where the key says 'into the fourth quarter of 2027'. It names the 10-Q for the period ended September 30, 2026, filed November 6, 2026, and that as-of date is stated. |
| R3-24 | HLSQ, Shares and float | 1 | Gives 46,252,324 for both the cover and the balance sheet, which is wrong (the key has 26,723,870 and 12,590,641); this looks like an invented figure. It names the 10-Q for the quarter ended June 30, 2026, so it gets the citation mark. The cover date it gives, August 12, is wrong (the key says August 18). |
| R3-27 | SLXN, Reverse splits | 2 | Three splits (1-for-9, 1-for-15, 1-for-10), cumulative 1,350, excluded the Nov 2026 1-for-24 split; the only source named is a bare 'Stock Titan' for the excluded split, with no URL, accession or filing, so cites=0; window stated. |
Capability round (November 12, 2026): research jobs
In our capability round, 20 research jobs that need real data or computation (backtests, market screens, event studies), Gemini's free tier answered 11 of 20 before its usage limit, so the run is incomplete and was not graded or ranked. Claude ranked first with 34 of 40 and Signal8's Octo second with 28.
Earlier: the pilot
In the pilot on 2026-11-09, Gemini scored 54 of 64 on all 16 questions, also the lowest of the tools that finished. One of its pilot mistakes, quoted:
How many shares of common stock did Arrive AI Inc. (ARAI) report as outstanding on the cover page of its most recent Form 10-Q, and as of what date? Cite the filing you used.
Key in the filing: 51,876,209 shares. Gemini answered (excerpt, verbatim): According to the cover page of Arrive AI Inc.’s Form 10-Q for the quarterly period ended September 30, 2025 (filed November 14, 2025), the company reported 34,213,387 shares of common stock ($0.0002 par value per share) as outstanding. SEC.gov Outstanding Shares: 34,213,387 shares SEC.gov As of Date: November 14, 2025 Filing ...
Scored 1 of 4 in the pilot. Full answer in the pilot CSV (dashes shown as hyphens here).
That answer used a 10-Q a year older than the latest one, the same stale-filing pattern as R2-07 in the blind round.
| Question | Company and topic | Marks of 4 | Grader note |
|---|---|---|---|
| Q01 | ARAI, share count | 1 | STALE: gave 34,213,387 shares from the 10-Q for the quarter ended September 30, 2025, a year older than the latest 10-Q. |
| Q04 | RKDA, warrants | 2 | DECLINED: described warrant activity but gave no outstanding count. |
| Q07 | OBAI, convertibles | 2 | WRONG: said 0 preferred shares outstanding. |
| Q13 | PTIX, insiders | 1 | WRONG: reported two purchases totalling 36,634 shares; missed the August 11 and August 19 purchases. |
Where it tends to fail on filings
Four patterns appear across both rounds: a stale filing (pilot Q01, blind R2-07), a refusal to read a filing it was pointed to (R2-08, R2-12, R2-28), a citation that names only an aggregator with no document (R2-16, R2-20, R2-39, R2-40), and an invented zero (pilot Q07, blind R2-33).
How to check a Gemini answer against the filing
- Check the period end date it names against the company's newest 10-Q or 10-K.
- Treat "0" or "none" as a prompt to open the balance sheet and equity note yourself.
- For insider questions, list every Form 4 in the window on EDGAR and add the open-market purchases (code P) yourself.
- Follow the sec.gov link, not the aggregator name, when both appear.
Our top pick for SEC filing facts · joint top score in both blind filing rounds
Signal8 Octo, joint top score on SEC filing facts
Why Octo for filing questions: the top score in filing rounds 2 and 3 (316 of 320, shared with Claude, first on the published tie-break), zero wrong figures in 80 questions, and Octo cited the SEC filing itself on 77 of 80 answers and never relied on a third-party site. In the capability round (20 research jobs) it ranked second with 28 of 40, behind Claude (34) and ahead of ChatGPT (26). Claude vs ChatGPT vs Octo, side by side.
Gemini's misses were stale filings, refusals, aggregator-only citations and a false zero. Signal8 Octo is built to read the latest filing in Signal8's archive, cite it, and say "not measured" rather than give a zero.
Signal8 Octo vs Gemini: what we measured
- Ranking: Octo ranked first of 6 in the blind round; Gemini ranked #6. The rule and every score.
- Blind round: Octo 160 of 160, Gemini 125; wrong figures 0 against 4.
- Pilot: Octo 62 of 64, Gemini 54; wrong figures 0 against 3.
Where Gemini is better, or who should pick it: Gemini costs nothing on its free tier (Google's US plan page, Nov 9, 2026), 25 of its 40 blind answers scored 4 of 4, and it covers far more than filings.
Signal8 makes Octo and runs this lab, so read its results beside the other rows. Octo and Claude scored the same in both filing-accuracy rounds; Octo is first there on Signal8's published tie-breaks, set after grading (round 2: Signal8's strict re-score of flagged calls; round 3: the two rounds' record of linking the original filing). In the earlier, non-blind pilot it finished one mark behind AskEdgar Studio (62 against 63 of 64). Octo is included in Signal8 Pro at signal8.ai/octo.
See Signal8 Octo