Claude for SEC filing research
Blind rounds 2026-11-10 and 2026-11-11: level with Octo, #2 on the tie-break in both · capability round November 12, 2026: #1 of five (corrected November 13, 2026) · plans checked Nov 9, 2026
Claude ranked first of the five complete runs in our capability round, 34 of 40 on 20 research jobs that need real data or computation, against 28 for Signal8's Octo and 26 for ChatGPT's free plan. That score was corrected on November 13, 2026: its first run scored 16, because our claude.ai account was left in Manual approval mode with network access off, so 9 tasks stopped at approval prompts no one answered. We re-ran those tasks on the same account and they were graded blind (correction). On SEC filing facts it is level with Octo on score. Claude answered all 40 questions in our blind test of SEC filing questions with full marks, 160 of 160, the top score, which it shares with Signal8's Octo. It ranks #2 of 6: level with Octo on score and on wrong figures (none), second on our strict-reading tie-break (about 155 against 159), which is Signal8's own re-score of flagged calls. In the harder round 3 it scored 156 of 160, again level with Octo and again with no wrong figure, and again ranks #2: the tie is broken by the two rounds' record of linking the original SEC filing, where Octo has 77 of 80 and Claude 63. One caveat weighs on this result more than on any other: the blind graders were themselves Claude instances, so we cannot rule out that they read Claude's answers more kindly.
What it is, and what it costs
Claude is Anthropic's assistant. Its pricing page (claude.com/pricing, read Nov 9, 2026) lists Free at $0 with "limited usage" every five hours and every week, web search included and Research not included. Pro is $20 a month billed monthly, or $17 a month on an annual plan ($200 billed up front), with more usage than Free and Research included. Max is $100 a month with "5x Pro usage" and Ultra is $200 a month. Team and Enterprise seats are also listed. Usage limits depend on the model, chat length and file size, which is why a long run of filing questions can hit the cap.
What we measured in the blind round
We ran Claude Opus 5.5 with web search switched on and with Signal8's own connectors (Signal8, DilutionScan and SplitScanner) switched off, so it could not read Signal8's data. Two early attempts made before the connectors were off were discarded and asked again, as was one attempt cut off by a usage limit. Every one of its 40 answers scored 4 of 4: right figure, a cited source the grader confirmed, and the right as-of date.
Every answer it gave in the blind round scored 4 of 4, so there is no lost mark to list.
How we rank: by blind score; equal scores are separated by (1) fewer wrong figures, then (2) the strict-reading score. Octo and Claude both scored 160 of 160 with no wrong figure, so the second tie-break decides first place: Octo 159, Claude about 155. The strict reading is Signal8's own re-score of the calls the blind graders flagged as lenient, not a mark the blind graders gave, and we set this rule on November 11, after grading. Round one's rule, fixed in advance on the methodology page, would report totals within 2 marks as level.
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. Claude scored 156 of 160 and was ranked #2 of the 4 tools that answered every question; Signal8's Octo ranked first with 156 and no wrong figure. Figure correct 40 of 40, cites a source 40, source contains it 36, as-of date right 40; wrong figures 0. Across rounds 2 and 3 it linked the original SEC filing on 63 of 80 answers, against 77 of 80 for Octo.
| Question | Company and topic | Marks of 4 | Blind grader's note |
|---|---|---|---|
| R3-12 | IVF, Reverse splits | 3 | 4 splits, 1-for-1,440; cites key 8-Ks for splits 2-4, but split 1 (1-for-12) rests only on a Nasdaq/RTTNews article that returned 403/timeouts. Checked: the cited July 2025 ex99-1 does not mention the March split, so split 1 is unverified and source = 0. |
| R3-16 | VTGN, ATM, warrants or shelf | 3 | Gave $175M total / $139.5M remaining at June 30, 2026, citing Panabee, Quantisnow (June 2025 424B5) and Webull; WebFetched Panabee and Webull and neither states $139.5M or the agreement size, and the 2025 424B5 cannot state a June 2026 balance. |
| R3-27 | SLXN, Reverse splits | 3 | Gave 3 splits 1-for-9/1-for-15/1-for-10, cumulative 1,350, excluded the Nov 2026 1-for-24; none of the cited docs is a key filing; sec.gov 424B3 confirms 1-for-9 and Finviz 1-for-10, but the Quantisnow page cited for 1-for-15 does not state it and no cited source states the full set. |
| R3-30 | CISO, Nasdaq status (8-K) | 3 | Gave Dec 28, 2026 as of Oct 31, 2026; cited edgar.tools (403, could not open), Boardroom Alpha Jan 2, 2026 8-K (first-period notice, June 29 deadline) and Quantisnow July 6 8-K page (fetched: headline only, no deadline stated); no key filing named. |
Capability round (November 12, 2026): research jobs
The capability round asked each tool to do 20 research jobs that need real data or computation: backtests, market screens, structured lookups, event studies and multi-step research. Claude scored 34 of 40 and ranked #1 of the five complete runs (14 done, 6 partial, 0 not done, none invented). Signal8's Octo was second with 28. This score was corrected on November 13, 2026. Claude's first run scored 16, because our claude.ai account was left in Manual approval mode with network access off, so 9 tasks stopped at a "needs your intervention" prompt that no one answered (C-02, C-06, C-07, C-09, C-10, C-14, C-15, C-16, C-17). We re-ran those tasks on the same account with Auto approval and network access on, and fresh graders graded them blind: all 9 done. The correction. Every grade.
Tie-break used: Octo and Claude both scored 156 of 160 in round 3 with no wrong figure, and 316 of 320 over both blind rounds. No strict re-score was made for round 3, so the tie is broken by the two rounds' combined record: wrong figures across all 80 questions (0 each, still level), then answers that linked the original SEC filing (Octo 77 of 80, Claude 63 of 80). This is Signal8's rule, chosen after round 3 was graded; Octo did not outscore Claude.
Where Claude's answers came from
Claude reached the right figure on every question in both blind rounds, often by way of other websites. It linked the original SEC filing on 63 of 80 answers, linked other websites (news, aggregators such as Stock Titan, company pages) on 59, and on 17 answers linked other websites only. Octo cited the SEC filing itself on 77 of 80 answers and never relied on a third-party site.
| Tool | Answers counted | Linked the original SEC filing | Linked another website | Other websites only | Round 2 / round 3: SEC filing linked |
|---|---|---|---|---|---|
| Signal8 Octo Made by Signal8, which runs this lab | 80 | 77 of 80 | 0 of 80 | 0 of 80 | 38 of 40 / 39 of 40 |
| Claude | 80 | 63 of 80 | 59 of 80 | 17 of 80 | 29 of 40 / 34 of 40 |
| ChatGPT | 80 | 50 of 80 | 12 of 80 | 6 of 80 | 31 of 40 / 19 of 40 |
| AskEdgar Studio | 80 | 26 of 80 | 1 of 80 | 0 of 80 | 16 of 40 / 10 of 40 |
| Perplexity | 79 | 51 of 79 | 27 of 79 | 16 of 79 | 24 of 40 / 27 of 39 |
| Gemini | Not measured | Not measured | Not measured | Not measured | Not measured |
Gemini is not measured: its links could not be captured from its page, so we report no figure rather than a zero. Perplexity counts 79 answers because one round 3 answer was not captured. Linking another website is not wrong in itself; it means part of the answer rests on a page that is not the filing. Per-answer codes: citation-sources.json.
Earlier: the pilot
Claude was not run in the pilot on 2026-11-09: the account we used had reached its usage limit, and on Nov 9, 2026 Claude still showed "Usage limit reached" with a reset at 4:00 AM on November 10. We did not work around the limit, so all 16 pilot rows for Claude read "not run".
How to check a Claude answer against the filing
- Ask it to quote the sentence from the filing, then find that exact sentence in the document on EDGAR.
- Ask which filing and which date it used, and compare with the newest filing on the company's EDGAR page; an answer built from an older filing is a common error in our data (Gemini did it in both rounds).
- If it says it searched the web, open the pages it cites and check each is the company's own filing.
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.
Signal8 Octo and Claude scored the same in both filing-accuracy rounds (316 of 320 each), so we do not claim Octo is more accurate than Claude on filing facts; Octo ranks first there on our published tie-breaks. In the capability round Claude ranked first with 34 of 40 and Octo second with 28. On filing facts the difference is in how they get there: Claude searched the web on a general assistant plan, while Octo answers from Signal8's archive of SEC filings and cites the filing for each answer.
Signal8 Octo vs Claude: what we measured
- Ranking: Octo ranked first of 6 in the blind round; Claude ranked #2. The rule and every score.
- Blind round 2: 160 of 160 each, no wrong figure from either; strict reading 159 against about 155.
- Blind round 3: 156 of 160 each, no wrong figure from either; linked the original filing on 77 against 63 of 80 answers over both rounds.
- Capability round: Claude 34 of 40, Octo 28. Claude's first run scored 16 because of a harness fault on our side; 9 tasks were re-run.
- Pilot: no comparison, because Claude was not run.
Where Claude is better, or who should pick it: Claude matched the top score on filing facts, ranked first on research jobs, and handles everything outside 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