Shared research report

What are the most significant developments in AI today?

November 21, 2026

Research Report

Question: What are the most significant developments in AI today?

Date: 2026-11-21T20:27:27.657866055+00:00

Coverage window: 2026-11-21 – 2026-11-21

Rounds: 4

Status: COMPLETE

What this run changed — 0 of 1 calls succeeded

Taken from the run's tool traces, not from its narrative: these calls actually went out.

Evidence: 85 claims · 57 sourced · 4 partial · 0 unsupported · 21 self-reported (no independent source) · 11 single-source

Executive Summary

The only development on 2026-11-21 that matters is the fallout from a claim made the night before: that an Anthropic mathematician, Levent Alpöge, used Claude to prove the Riemann hypothesis — a Millennium Prize problem — and that a Lean formalization of it now passes machine verification. A verification page last updated 2026-11-21 12:55 UTC reports a leanchecker kernel replay of 175,187 of 175,187 checks with zero rejections (math.mit.edu/~drew/grh). Le Monde published a same-day science-desk analysis on 2026-11-21 (lemonde.fr) and The Guardian reported the claim on 2026-11-21, noting verification was still underway (theguardian.com). The claim itself was posted Friday 20 November, so today is coverage and formal checking of a one-day-old announcement, not a fresh discovery.

Everything else being sold as "today's AI news" is a 4–20 November event re-dated. Google's Genie 4 API access (19 Nov), Gemini 4 Neon (18 Nov) and Gemini Tasks phone calling (17 Nov), Tencent's Hy4 weights (20 Nov), Anthropic's $81.34B IPO (priced 18 Nov), and OpenAI's agent benchmark (18 Nov) are all outside the window on the labs' own records. No funding round, chip announcement, datacenter deal, regulator action or safety-incident disclosure verifiably dated 2026-11-21 exists in the record gathered here. It was a Saturday; Nasdaq was closed and every frontier lab newsroom checked was silent.

What Actually Happened on 2026-11-21

ItemWhat it isDate evidenceStatus
Riemann claim — formal checkingleanchecker kernel replay: 175,187/175,187, zero rejections; page states its own scope limitsPage last updated 2026-11-21 12:55 UTC (link)Verified in-window; verification incomplete
The Guardian reportAlpöge posted a note Friday saying Claude helped prove the Riemann hypothesis over ~6 weeks; linked Lean code; mathematicians "began working to see if Alpöge's proof could be confirmed"Byline Sat 21 Nov 2026 11:39 EST, modified 15:01 EST (link)Verified in-window publication; event is 20 Nov
Le Monde analysisOriginal staff piece by David Larousserie; states the problem was solved "le 20 novembre"; paywalled at 83% remainingdatePublished 2026-11-21T10:30+01:00 (link)Verified in-window publication; not a new event
Inference-mix reportingPositron AI at SC26: ~two-thirds of workloads are inference vs one-third training (2023: 85/15); Oracle bringing ~340 Atlas servers live this quarter; 4,500W/server; 40kW per Titan rack; 90–120 days land-and-power to tokensPage "UPDATED 13:28 EST / NOVEMBER 21 2026" (link)In-window journalism; single vendor exec on a sponsored segment; not confirmed at Oracle or Positron primary source
Consumer-product itemsTechCrunch's day file: "top AI agents that live in your text messages"; a microdrama startup buying AI likeness rightsTechCrunch 2026/11/21 archive (link)In-window; no funding, capacity or silicon figures

The Day's Real Story: A Claim, Not a Result

Alpöge's own words, as quoted by The Guardian: "Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line." On timing: "The precise timing is horrible. If I could have, I'd have finished this all much earlier" — the note landed one day after Anthropic's Nasdaq debut.

Three things make this contested rather than settled:

What Was Recycled as "Today's News"

Item circulated on 21 NovActual primary datePrimary evidence
Genie 4 landed in the Gemini API2026-11-19datePublished 2026-11-19T17:00Z; post opens "Following yesterday's Genie 4 launch" (blog.google)
Genie 4 / Project Genie availability2026-10-27 announced; 2026-11-18 consumer rolloutdatePublished 2026-10-27T16:00Z and 2026-11-18T18:00Z (blog.google)
Gemini 4 Neon launch2026-11-18Model card header "Published 18 November 2026" (deepmind.google); blog JSON-LD 2026-11-18T21:00Z
Gemini Tasks can phone businesses2026-11-17datePublished 2026-11-17T17:00Z, byline "Nov 17, 2026"; six countries: US, Japan, Korea, Canada, Australia, India (blog.google)
Tencent Hy4 open weightsweights pushed 2026-11-20T03:05:49Z; repo created 2026-10-29; press release 4 NovHF API lastModified/createdAt (huggingface.co/tencent/Hy4)
Tencent Hy4-flashfirst commit 2026-11-17, last write 2026-11-19HF commits API (link)
Anthropic IPOpriced 2026-11-18, traded 19 Nov, settled 20 NovForm FWP: 980,000,000 Class A shares at $83.00; $81.34B excluding greenshoe; $26.0B concurrent Class N placement (sec.gov); newsroom close release 20 Nov, 1,127,000,000 shares incl. 147,000,000 option shares (anthropic.com/news)
OpenAI agents solve 46% of 2–4h tasks2026-11-18 reportOpenAI newsroom lists "Research acceleration: The jagged frontier — Nov 18, 2026" (openai.com/news); TNW's 21 Nov piece is coverage
"AI incidents on 21 Nov" in the AIIDIncidents dated 17 and 18 Nov; 21 Nov is date_downloadedIncident pages 1835/1836 (incidentdatabase.ai/cite/1836)

Two figures should be corrected rather than repeated. "$1.5 trillion" for Anthropic is not in the pricing document — it gives a raise size, not a market cap; the only observable marks are a 19 Nov close implying ~$1.28T and a 20 Nov close of $88.91 implying ~$1.17T (both derived). "774B" for Hy4 is not a contradiction of "770B": Tencent's card says 770B total / 49B active for the language backbone, and 774,434,002,142 is Hugging Face's all-in safetensors count including the vision encoder, adapter and drafter.

The Empty Verticals

Lab-by-Lab

Lab2026-11-21 statusBasis
AnthropicNot a corporate announcement; third-party coverage of an employee's claimNewsroom newest item: IPO close, 20 Nov (link)
OpenAINothing publishedNewest index item: MedChemBench, 20 Nov (link)
Google / DeepMindNothing published on 21 Nov; its 17–19 Nov cluster is being recycleddatePublished fields on blog.google posts; deepmind.google RSS newest 18 Nov
xAINothing publishedx.ai/news newest dated item 16 Nov (link)
MistralNothing publishedRSS newest item 5 Nov (link)
CohereNothing publishedNewest index item 17 Nov (cohere.com/blog)
TencentNo 21 Nov release; Hy4 landed 20 NovHF lastModified
Meta AI, DeepSeek, Qwen, Baidu, ByteDanceUnverified — not checkedBlog/newsroom fetches did not complete

Analysis

The shape of this day is explained by three things, and none of them is "a slow news day."

A genuinely large event landed on Friday. A claimed AI-assisted proof of a Millennium Prize problem is the kind of story that generates 48 hours of coverage; Saturday's output is that coverage, plus the first machine-verification artefact. The window therefore contains a real, checkable development — just not a new announcement.

Daily roundup sites are re-dating last week's cluster. Every mis-dated item above traced to the same failure mode: a same-day "AI news today" page leading with a product that Google, Tencent or OpenAI date days earlier on its own structured metadata. One such page led with Gemini 4 Neon and its 61-on-Artificial-Analysis score — a score that appears in Google's 18 November post. Where a primary datePublished field exists, it overrides the roundup; treat these sites as link indexes, never as evidence.

Saturday suppresses the categories that usually fill this brief. No Nasdaq session, no funding announcements, no regulator filings. The absence of capital, policy and safety items is consistent with the calendar rather than proof of quiet, but it is the fourth independent sweep to come back empty on the same verticals.

Risks & Open Questions

Claims without independent support

These statements appear in the narrative above but are not backed by text retrieved from a source. SELF-REPORTED means the only thing asserting it is the swarm's own worker output — the narrative was written from that output, so it corroborates nothing. Treat all of these as unverified.

Bibliography

  1. math.mit.edu/~drew/grh
  2. lemonde.fr
  3. theguardian.com
  4. link
  5. link
  6. blog.google
  7. blog.google
  8. deepmind.google
  9. blog.google
  10. huggingface.co/tencent/Hy4
  11. link
  12. sec.gov
  13. anthropic.com/news
  14. openai.com/news
  15. incidentdatabase.ai/cite/1836
  16. aienews.org/digest/2026-11-21
  17. POLITICO via MSN
  18. MSN
  19. MSN
  20. link
  21. link
  22. link
  23. cohere.com/blog

Detailed Findings

Round 0 · Finding 1

AI Model, Product & Feature Releases — 2026-11-21

Executive Summary

I could not find a single high-authority primary announcement (an OpenAI/Google/Anthropic/Meta newsroom post) that I was able to open and date-stamp to 2026-11-21 within my fetch budget. What the live web does show for this date is a cluster of daily AI roundup/aggregator pages published on 2026-11-21, plus two primary artifacts I did open (Google DeepMind's Genie 4 model page and Tencent's Hugging Face organisation page). Those primary artifacts confirm the existence of the products but do not themselves carry a visible 2026-11-21 date, and in Tencent's case the Hugging Face timestamps point to ~2026-11-19 rather than 2026-11-21.

So the honest answer to "what shipped on 2026-11-21" is: a Google Gemini/Genie release cluster and a Tencent open-weights release are the headline technical items reported for that day, but both rest substantially on secondary coverage, and one of them (Tencent) is contradicted by the primary registry's own timestamps. Confidence in the date-pinning is therefore low-to-moderate.

Overall confidence: 0.45 — the items are real, the exact-day attribution is not solidly established.


Key Findings

1. Google — Genie 4 made available in the Gemini API / AI Studio (CONFIDENCE: MEDIUM that it exists; LOW that 2026-11-21 is the availability date)

2. Google — Gemini "Tasks" can place phone calls to businesses on the user's behalf (CONFIDENCE: LOW — secondary only)

…(truncated — the summary above captures the substance)

Round 0 · Finding 2

AI Developments Dated 2026-11-21 — Research, Safety & Controversy Focus

Scope note on method. I treated the "daily roundup" sites that dominate search results for this date (aienews.org, aitoolsrecap.com, thirdruntime.com, pasqualepillitteri.it, artificialintelligenceherald.com, creati.ai) as indexes only, not as evidence. Several are machine-generated aggregators (aitoolsrecap.com's own schema gives a foundingDate of 2026-03-18). Every claim below is anchored either to a page I fetched directly, or is explicitly labelled as headline-level/unverified. I also pulled the primary newsroom indexes of OpenAI and Anthropic to test whether same-day primary announcements existed.


1. Executive Summary

The single most consequential item carrying a verifiable 2026-11-21 publication date is a research/priority dispute in mathematics: a Guardian report, timestamped 21 Nov 2026, that an Anthropic mathematician claims Claude helped him prove the Riemann hypothesis — with the claim itself posted the previous day and not yet independently verified.

The rest of the day is thinner than the aggregators suggest. Two of the loudest "Nov 21" AI stories are coverage of announcements made earlier in the week (OpenAI's research-acceleration report, dated 18 Nov on OpenAI's own newsroom). On the primary record, neither OpenAI nor Anthropic published a new item dated 21 Nov 2026 in the newsroom listings I fetched. A genuine same-day-dated compute/infrastructure story exists (SiliconANGLE, updated 13:28 EST 21 Nov). Policy items dated to the day exist but only at headline level — I could not open the underlying articles (paywall / not fetched), so they are reported as UNVERIFIED.

No primary AI safety-incident report or evaluation release dated 2026-11-21 was verified. I state that as a gap rather than filling it.


2. Key Findings (with confidence levels)

…(truncated — the summary above captures the substance)

Round 0 · Finding 3

AI Policy & Legal Actions — 2026-11-21 (single-day window)

Executive Summary

I ran this as a strict single-day window search (2026-11-21 only) and the in-window, policy/legal yield is thin and, critically, entirely second-hand. My general web-search path degraded badly — search_engine_results on both "auto" and "duckduckgo" returned generic encyclopedia/marketing pages for every AI-policy query I tried, Google News RSS returned an empty body, and apnews.com was hard-blocked by bot protection. The one surface that did return live, timestamped AI news was Bing News' 7-day filter, fetched on 2026-11-21, which surfaced a short list of same-day items.

Bottom line: I found no primary-source government record (agency docket, court filing, gazette, official press release) confirming any AI policy or legal action on 2026-11-21. What I found is a small number of news reports published within hours of my fetch. I am reporting them as reported, not as confirmed, and flagging the date provenance on each. I did not find enough verified material to meet a "3 distinct confirmed developments" bar for this facet.


Key Findings

1. California Attorney General signals independent antitrust scrutiny of AI (jurisdiction: California, USA) — MODERATE-LOW confidence

…(truncated — the summary above captures the substance)

Round 0 · Finding 4

AI Developments Dated 2026-11-21 — Capital, Compute, and Same-Day News

Research cut-off: 2026-11-21 ≈20:15 GMT. Important framing: the requested date is today, and the day is only partially elapsed at cut-off. I report what dated sources actually show, and I state plainly where the requested category is empty.


1. Executive Summary

The specific category requested — funding rounds, acquisitions, Nvidia/AMD/custom-silicon announcements, and datacenter or power deals dated 2026-11-21 — returned no verified in-window items. I could not find a single same-day funding round, chip announcement, or datacenter/power deal carrying a dollar or capacity figure with a 2026-11-21 publication date. This is a genuine negative finding, not an omission, but it is bounded: the day was ~20 hours old at cut-off and I could only read the first portion of the day's aggregated feed (limitation detailed in §5).

What is verifiably dated 2026-11-21 is thinner and more diffuse than a typical news day: one significant AI-research claim reaching a major outlet, one peer-reviewed-adjacent science result, and a set of trade/vertical and commentary pieces. The largest AI business stories in circulation — Anthropic's IPO, the Jensen Huang exit from Nvidia, US chip-export legislation — are all pre-window (≈Nov 15–19) and appear in today's feeds only as background or follow-up.

Headline conclusion: for 2026-11-21, the economic dimension of the AI story is absent from primary same-day reporting; the day's AI news is research- and product-adjacent, and the consequential capital/compute events belong to the preceding week.


2. Key Findings (with confidence)

#FindingDate basisConfidence
1No funding round, acquisition, chip, or datacenter/power deal dated 2026-11-21 foundDirect inspection of TechCrunch's 2026/11/21 day archive + Google News 24h AI/chip feedMedium-High (bounded by partial day + partial feed read)
2NYT published a same-day story on an Anthropic mathematician's Riemann Hypothesis claim using ClaudepubDate Sat, 21 Nov 2026 16:52:00 GMTHigh that the article is dated 11/21; Low on underlying substance (unverified)
3ScienceDaily reported a study finding AI does not work like the human brainpubDate Sat, 21 Nov 2026 15:25:56 GMTMedium (headline verified; underlying paper not verified)
4The Riemann claim itself surfaced ~2026-11-20, making the 11/21 NYT piece follow-up coverageTechCrunch "1 day ago" relative timestampMedium-High
5No same-day item found for OpenAI, Google/DeepMind, Meta, xAI, Microsoft, or NvidiaAll fetched indexesMedium (absence of evidence)
6No policy/regulatory/legal action dated 2026-11-21 with a named issuing body verifiedSameMedium

3. Detailed Analysis

3.1 The capital-and-compute category is empty for 2026-11-21

…(truncated — the summary above captures the substance)

Round 1 · Finding 1

Verdict

No. Neither tencent/Hy4 nor tencent/Hy4-flash shipped on 2026-11-21. Both repos existed, and carried weights, before the window opened. The only Tencent Hugging Face activity I can date to 2026-11-21 is a set of weight/quantisation updates to the WeMM-Video family — which is consistent with, but does not by itself prove, the claim that the same-day Tencent item was WeMM-Video-35BA3B.

The prior-round contradiction resolves against "774B Hy4 shipped this week." Hy4 is roughly a three-week-old repo whose weights landed on 11-20, and the "774B" figure is itself not what Tencent's own card says.


Key findings

1. tencent/Hy4 — repo created 2026-10-29, weights pushed 2026-11-20. Not an 11-21 release. (Confidence: high)

FieldValueSource
createdAt2026-10-29T08:29:16Zhttps://huggingface.co/api/models/tencent/Hy4?expand[]=createdAt
Commit count1 (x-total-count: 1)https://huggingface.co/api/models/tencent/Hy4/commits/main
Sole commitc10c0de "update", author yiqichen01, 2026-11-20T03:05:49Zsame
lastModified2026-11-20T03:05:49Zhttps://huggingface.co/api/models/tencent/Hy4

The repo is 23 days older than the window and its single commit predates the window by a day. Its lastModified (2026-11-20) is the most recent write, and it is outside 2026-11-21..2026-11-21.

The "774B" figure is wrong at the primary source. Tencent's own model card specifies 770B total parameters ("The language backbone comprises 770B total parameters, of which 49B are activated per token"), with a 1.23B ViT and a 3.3B DFly2 drafter — so a secondary "774B" total is a plausible rounding-up of 770B + 3.3B drafter, not a Tencent-stated number. Model card: https://huggingface.co/tencent/Hy4 (spec table: "Total Parameters (excluding ViT, adapter, and DFly2) — 770B"; "Activated Parameters per Token — Text 49B + Vision 1.23B").

2. tencent/Hy4-flash — first commit 2026-11-17, last write 2026-11-19. Not an 11-21 release. (Confidence: high)

FieldValueSource
Commit count7https://huggingface.co/api/models/tencent/Hy4-flash/commits/main
Earliest commitbf68d9f "update", 2026-11-17T13:19:30Zsame
Latest commitb2498eb "update", 2026-11-19T11:25:34Zsame
lastModified2026-11-19T11:25:34Zhttps://huggingface.co/api/models/tencent/Hy4-flash

createdAt was not returned directly (my expand call for it errored, and the org listing below cut off before this repo). It is bounded: the earliest commit is 2026-11-17T13:19:30Z, and the repo does not appear in the top-20 most-recently-created Tencent repos, whose 20th entry has createdAt 2026-11-17T08:55:11Z — so createdAt ≤ 2026-11-17T08:55:11Z. Either way it is well before 2026-11-21. I am stating the bound rather than a fabricated exact timestamp.

…(truncated — the summary above captures the substance)

Round 1 · Finding 2

Bottom line

Google did not publish anything about Genie 4 or Gemini Tasks on 2026-11-21. Every Google-owned record I reached for these two items carries a publication date of 2026-10-27, 2026-11-17, 2026-11-18, or 2026-11-19 — all before the window. The "2026-11-21 Genie 4 / Gemini Tasks" framing in the secondary snippets is a date error, not a same-day launch. Separately, I could not recover the five excluded US states from Google's own post: that footnote fell inside the truncated portion of the page, so I am reporting it as unverified rather than filling it from secondary reporting.


1. Genie 4 — dates as recorded on Google-owned properties

Three distinct Google publications, three distinct dates, none in-window:

Google propertyHeadlineDate stamped on the page
blog.google"We're announcing Genie 4, our new world model for real-time interactive simulation."Oct 27, 2026
blog.google"Genie 4 is now available" (Project Genie rollout to AI Ultra)Nov 18, 2026
blog.google"Genie 4 is now available on the Gemini API."Nov 19, 2026

Evidence: the announcement post's JSON-LD reads "datePublished": "2026-10-27T16:00:00+00:00" and the body prints "Oct 27, 2026" (https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/). The subscriber-rollout post reads "datePublished": "2026-11-18T18:00:00+00:00", body "Nov 18, 2026" (https://blog.google/innovation-and-ai/models-and-research/google-deepmind/genie-4-now-available-project-genie/). The developer post reads "datePublished": "2026-11-19T17:00:00+00:00", body "Nov 19, 2026", and opens "Following yesterday's Genie 4 launch" — which independently corroborates the Nov 18 date (https://blog.google/innovation-and-ai/technology/developers-tools/genie-4-in-gemini-api/).

Confidence: high. These are machine-readable datePublished fields emitted by Google's own CMS, not prose.

Two consequences worth stating plainly:

Snippet-only, not verified by fetch: Google DeepMind's Genie 4 model card at https://deepmind.google/models/model-cards/genie-4/ and the Genie 4 research page at https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ (the latter is linked from both blog posts but I did not open it). Bing surfaced the model card as "3 days ago," i.e. ~Nov 18 — treat that as a snippet timestamp, not a confirmed date.

2. Gemini Tasks — the launch post is Nov 17, not Nov 21

…(truncated — the summary above captures the substance)

Round 1 · Finding 3

Did Anthropic announce a Riemann-hypothesis solution on 2026-11-21?

Executive Summary

The claim is substantively real but mis-dated, and it is not an Anthropic announcement. Every dated source I could retrieve places the disclosure on 2026-11-20, not 2026-11-21 — a one-day miss against this task's single-day window. The claim was made by an Anthropic employee (Levent Alpöge) in a personal essay, not by Anthropic as a company: I found no Riemann-hypothesis post on Anthropic's newsroom or research index. A competing, explicitly non-AI proof of the Generalized Riemann Hypothesis was released the same day (Nov 20), not on Nov 21, by Stanford postdoc Alessandro Masullo via AMS president Ravi Vakil.

Net: the story is a same-day AI-vs-human formalization race dated Nov 20, not an Anthropic corporate announcement on Nov 21. Several artefacts the task asks me to check (arXiv ID, GitHub/Lean repo, Zulip thread, Tao commentary, Quanta, MathOverflow) I could not reach before my tool budget expired — those are named as unverified below, not asserted.


Key Findings (with confidence)

#FindingConfidence
1The Riemann claim was disclosed 2026-11-20, not 2026-11-21High
2It was not an Anthropic corporate announcement — no such post on anthropic.com/news or /researchMedium-High (one unread hero headline — see gap)
3A competing all-human proof (of GRH) was released the same day, Nov 20 — not Nov 21Medium-High
4arXiv ID / GitHub repo / Lean identifier: NOT VERIFIED — I could not locate onen/a (negative result)
5Expert/Zulip/Quanta/MathOverflow/Tao commentary: NOT VERIFIEDn/a (negative result)

Detailed Analysis

1. The date is November 20, not November 21

TechCrunch's article carries an explicit timestamp of 11:32 AM PST · November 20, 2026, and states Alpöge "wrote in an essay announcing his findings," having messaged Dario Amodei "last Sunday night" about how to announce a "Lean-compilable solution to the Riemann hypothesis." (https://techcrunch.com/2026/11/20/ai-claimed-a-solution-to-the-167-year-old-riemann-hypothesis-and-a-human-did-too/)

Independent corroboration of the Nov 20 date comes from URL date-stamps and datelines visible in search results, though I did not fetch these pages, so I mark them as snippet-level only:

…(truncated — the summary above captures the substance)

Round 1 · Finding 4

Verification Sweep: Five 2026-11-21 Claims Against Primary Records

Bottom line up front: None of the five claims is supported by a primary source dated 2026-11-21. One of them — the Anthropic IPO — turns out to be real but mis-dated: it is a 2026-11-18/19 event fully documented on SEC EDGAR, and the "$1.5 trillion" figure attached to it is not in the pricing document I read. The other four produced no primary-source hit at all in the searches I ran. Because my tool budget closed before I could reach every regulator press room directly, I am marking those four "not found at primary source" rather than "false" — the distinction matters and I flag exactly what I did not check.


1. Executive Summary

The single most important correction this sweep produces is that the Anthropic IPO is not a snippet-level rumor that evaporates — it is a completed, EDGAR-documented offering. Anthropic, PBC trades as ANT on Nasdaq (CIK 0002130357). But it priced on 2026-11-18, traded 2026-11-19, and settled 2026-11-20. EDGAR shows zero Anthropic filings dated 2026-11-20 or 2026-11-21. So the item cannot be reported as "today's development" for 2026-11-21 without mis-dating it.

The other four items (Oracle/Positron, California AG, Trump data-center-abroad tax, UC AI agreement) returned no relevant primary or even secondary result across four targeted searches. That is a weaker signal than a negative finding — it is an absence of evidence, and I name the specific properties I failed to reach.


2. Claim-by-Claim Verdicts

(a) Anthropic IPO / "$1.5 trillion" — PARTIALLY CONFIRMED, MIS-DATED, FIGURE UNSUPPORTED

SEC EDGAR result: PRESENT. This is the decisive finding.

EDGAR company search for Anthropic returns a full IPO docket — 32 filings: https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0002130357&type=&dateb=&owner=include&count=100

The docket, in reverse order (filing dates as stamped by EDGAR):

FormEDGAR filing dateAccession
424B4 (final prospectus)2026-11-190001628280-26-074944
FWP (pricing)2026-11-190001628280-26-074937
S-8 ×22026-11-190001628280-26-074934 / -074931
EFFECT (notice of effectiveness)2026-11-189999999995-26-003574
S-1MEF (Rule 462(b))2026-11-180001628280-26-074904
CERT (Nasdaq listing cert)2026-11-180001354457-26-001088
8-A12B (Section 12(b) registration)2026-11-170001628280-26-074342
S-1/A No. 32026-11-170001628280-26-074338
S-1/A No. 22026-11-100001628280-26-072484
S-1/A No. 12026-11-020001628280-26-068676
S-1 (initial)2026-10-260001628280-26-067168
DRS / DRS-A ×62026-06-01 → 2026-10-09File 377-09429

…(truncated — the summary above captures the substance)

Round 2 · Finding 1

Anthropic (NASDAQ: ANT) — IPO pricing, first-day trading, and the "$1.5 trillion" question

Executive Summary

Anthropic, PBC went public on the Nasdaq Global Select Market under the ticker ANT, priced at $83.00 per share on 18 November 2026, with the 424B4 final prospectus filed 19 November 2026. It is the largest IPO on record: 980,000,000 Class A shares for gross proceeds of $81.34 billion (FWP, dated 18 Nov 2026; EDGAR filing index, 424B4 filed 2026-11-19).

The "$1.5 trillion" figure is UNCONFIRMED as a market capitalisation and should not be reported as one. No primary document I fetched states any valuation at all — the pricing FWP is silent on it. The number is arithmetically consistent only with a fully diluted share count at the $83 offer price, not with the shares-in-issue market cap at any traded price. The observed market capitalisation was ~$1.28–1.3 trillion at the 19 Nov close and ~$1.17 trillion at the 20 Nov close.

There was no 21 November trading session. 2026-11-21 was a Saturday; Nasdaq's own historical-quote API returns rows for 11/19 and 11/20 only, with nothing for 11/21 (Nasdaq quote API). The last observable price in the window is the 20 Nov close of $88.91.

Key Findings (with confidence levels)

1. Pricing terms are confirmed at primary source. (Confidence: HIGH) The issuer free writing prospectus, dated 18 November 2026, states: 980,000,000 Class A shares; size excluding the option $81,340,000,000; price to public $83.00; price to underwriters $82.43878; base shares 956,351,354 primary plus 23,648,646 from selling stockholders; an option on up to 147,000,000 additional primary shares; a concurrent $26.0 billion private placement of Class N stock (313,253,007 shares) at the price to public; pricing date 18 Nov 2026, trade date 19 Nov 2026, settlement date 20 Nov 2026; CUSIP 03703U 107 (FWP). Arithmetic check: 980,000,000 × $83.00 = $81.34bn, matching the stated size exactly.

…(truncated — the summary above captures the substance)

Round 2 · Finding 2

Verification status of the 2026-11-20 Riemann-hypothesis claim, as of 2026-11-21

Bottom line: The claim is formalised and partially machine-verified, but not validated and not refuted as of 2026-11-21. It has not gone unaddressed — it drew heavy expert attention within 24 hours. The single most important in-window artefact is an MIT-run Lean verification page timestamped 2026-11-21 12:55 UTC that reports the kernel checks largely passing while two checks were still running.

1. What the claim actually is (and the conflation to avoid)

Two separate events landed on 2026-11-20 and are routinely merged in coverage:

The two claims are not the same result and should not be reported as one.

2. The in-window primary verification artefact (2026-11-21)

Andrew Sutherland (MIT) published "RH Lean proof: verification files", page last updated 2026-11-21 12:55 UTC, explicitly stating results are "as of that time and this page will be updated as the remaining checks finish." https://math.mit.edu/~drew/grh/

…(truncated — the summary above captures the substance)

Round 2 · Finding 3

Executive Summary

Tencent's Hy4 is a 770B-parameter (language backbone) sparse MoE multimodal flagship whose weights went live on Hugging Face at 2026-11-20T03:05:49Z — not 2026-11-21, and not 2026-11-19 in canonical UTC terms. The "774B" figure that appeared in earlier rounds is not a competing claim about the model; it is the all-in safetensors parameter count (774,434,002,142) published by Hugging Face's own metadata, which includes the vision encoder and the speculative-decoding drafter that Tencent explicitly excludes from its 770B headline. Both numbers are correct and they reconcile.

Critically for the 2026-11-21 window: there is no primary artefact dated 2026-11-21 for Hy4. The Hugging Face repo's single commit, lastModified, and the GitHub repo's final README commit all terminate on 2026-11-20. Hy4 is therefore a 2026-11-20 (UTC) event — an adjacent-window development, roughly 24–41 hours before the window closes — and cannot be reported as a 2026-11-21 development.


Key Findings

1. Parameter count: 770B is the primary-source headline figure; 774.4B is the primary-source all-in weight count. (Confidence: HIGH)

The official model card, fetched from both Hugging Face and GitHub, states verbatim:

"The language backbone comprises 770B total parameters, of which 49B are activated per token."

The specification table gives "Total Parameters (excluding ViT, adapter, and DFly2): 770B", with "ViT Parameters: 1.23B" and "DFly2 Parameters: 3.3B" listed as separate line items. — https://huggingface.co/tencent/Hy4/raw/main/README.md — https://github.com/Tencent-Hunyuan/Hy4

The Hugging Face model API returns the machine-computed tensor count:

"safetensors":{"parameters":{"BF16":774403206721,"F32":30795421},
"total":774434002142}

i.e. 774,434,002,142 parameters ≈ 774.4B, with usedStorage 1,548,935,130,170 bytes (≈1.55 TB, consistent with BF16 storage). — https://huggingface.co/api/models/tencent/Hy4?expand[]=safetensors&expand[]=createdAt&expand[]=lastModified&expand[]=usedStorage

Resolution of the 770B vs 774B dispute: these are the same model measured against two different denominators. Tencent's 770B excludes the ViT (1.23B), the adapter, and the DFly2 drafter (3.3B); Hugging Face's 774.4B counts every tensor in the repository. The gap between the two figures is accounted for by those separately-declared components. Neither figure is an error, and neither source contradicts the other. The 774B number should not be cited as "Tencent says 774B" — Tencent says 770B.

2. Weights publication date: 2026-11-20 (UTC). (Confidence: HIGH)

Three independent primary timestamps agree:

…(truncated — the summary above captures the substance)

Round 2 · Finding 4

2026-11-21 lab newsroom sweep — direct index checks

Scope of this sub-task: determine whether OpenAI, Google DeepMind, Anthropic, Meta AI, xAI, Mistral, Cohere or Tencent published anything timestamped 2026-11-21 (UTC), by reading each lab's own newsroom/blog index page rather than search snippets. I read six of the eight indices directly; two could not be read (details below). I did not reach AIID or any lab safety-incident page — that part of the question is unanswered here, not answered negatively.


Executive summary

Across the six lab indices I was able to read, no item carries a 2026-11-21 publication date. The newest dated entry on each index is 2026-11-20 or earlier. Two indices (DeepMind, Mistral) publish no day-level dates on the listing page at all, so they neither confirm nor exclude a 2026-11-21 post. Meta AI's index timed out on three attempts and Tencent was not reached, so those two are unchecked.

This is a negative result with a real evidentiary basis for four labs (OpenAI, Anthropic, xAI, Cohere) — each index is sorted newest-first, so a 2026-11-21 post would have appeared above the item I found at the top.


Index-by-index results

LabIndex URL fetchedNewest dated item on index2026-11-21 post?
OpenAIhttps://openai.com/news/MedChemBench: Benchmarking AI for medicinal chemistry — Research, Nov 20, 2026No
Google DeepMindhttps://deepmind.google/blog/Introducing Gemini 4 Neon — "November 2026" (no day)Undetermined
Anthropichttps://www.anthropic.com/news/ and https://www.anthropic.com/news/announcementsAnthropic announces closing of initial public offering… — Nov 20, 2026No
Meta AIhttps://ai.meta.com/blog/—Unchecked (timeout ×3)
xAIhttps://x.ai/newsBuilding an accounting team with Grok Bot — Product · Nov 16, 2026No
Mistralhttps://mistral.ai/news/none — index shows no dates on any of 94 resultsUndetermined
Coherehttps://cohere.com/blogCrunchbase in North: Private market signals for growth-stage companies — Nov 17, 2026No
Tencent (Hunyuan)not fetched—Unchecked

Findings with evidence

1. OpenAI — nothing on 2026-11-21. (High confidence) The news index is ordered newest-first and its top entry is MedChemBench: Benchmarking AI for medicinal chemistry, tagged Research, dated Nov 20, 2026 (https://openai.com/news/). Below it: How OpenAI's Performance Factory keeps ChatGPT fast (Nov 19), The Silent Consensus (Nov 18), Research acceleration: The jagged frontier (Nov 18), Inside the Codex sandbox (Nov 18), Managing frontier AI risk: From training to deployment (Nov 18), Le Groupe Figaro and OpenAI partner… (Nov 18), Advancing chip design with OpenAI models (Nov 18), Introducing GPT-6.2 Astra and Sol (Nov 17). No 2026-11-21 entry exists above the Nov 20 item.

…(truncated — the summary above captures the substance)

Round 3 · Finding 1

Sub-task answer: DeepMind "Gemini 4 Neon" publication date, and whether deepmind.google published anything on 2026-11-21

Bottom line

"Gemini 4 Neon" was published on 2026-11-18, not 2026-11-21. The "November 2026" label that the prior round could only recover is a month-level display label on DeepMind's model-card page; the same page carries an exact day-level stamp in its header. DeepMind published nothing dated 2026-11-21 — its RSS feed's newest item is dated 18 Nov 2026, and its blog index contains no September–November 2026 entry with a 21 November date.

This closes the ambiguity flagged as most likely to flip the answer: Gemini 4 Neon is not the 2026-11-21 headline. The day remains a media/incident story (outside this sub-task's scope).


1. Exact publication date of Gemini 4 Neon — 2026-11-18 (high confidence)

Four independent primary/structured signals, all on Google-controlled properties, agree:

SignalValueURL
Model card page header"Published 18 November 2026"https://deepmind.google/models/model-cards/gemini-4-neon/
blog.google JSON-LD datePublished2026-11-18T21:00:00+00:00https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
blog.google visible byline"Nov 18, 2026"same URL
DeepMind RSS <pubDate>Wed, 18 Nov 2026 21:35:43 +0000https://deepmind.google/blog/rss.xml
DeepMind sitemap lastmod for the model card2026-11-18https://deepmind.google/sitemap.xml

How the "November 2026" ambiguity arises, precisely: the same page renders both granularities. At the top it prints "Published 18 November 2026", and further down, in the body block, it prints "Published: November 2026" (https://deepmind.google/models/model-cards/gemini-4-neon/). A crawler or index that captured only the body label — or DeepMind's blog index cards, which show only "November 2026" — loses the day. The day-level value is recoverable from the page header, the RSS feed, and the sitemap.

Also worth noting: dateModified is 2026-11-19T14:03:59.958636+00:00 (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) — so the piece was edited on 19 Nov, still not the 21st. Anyone dating it by "last touched" also lands outside the window.

Correction to a possible misreading: https://deepmind.google/blog/introducing-gemini-4-neon/ does not host an article. On fetch it resolves with Canonical: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/ — DeepMind's blog path is a referral wrapper around blog.google. So "DeepMind's post" and "Google's post" are one document, published 2026-11-18T21:00Z.

…(truncated — the summary above captures the substance)

Round 3 · Finding 2

Did Mistral, Meta AI, Tencent Hunyuan, DeepSeek, Alibaba Qwen, Moonshot or Zhipu publish anything dated 2026-11-21?

Executive Summary

No. Across every lab I could reach through a dated primary channel (RSS feed, blog index with visible post dates, or official docs), I found zero items published on 2026-11-21. The single most consequential finding is a correction to the framing of the whole lab sweep: DeepMind's "Gemini 4 Neon" was published 2026-11-18, not 2026-11-21, and the "November 2026" label was an artifact of the DeepMind index page, not a date ambiguity in the post itself. Mistral's Atom feed proves the company's most recent post is 5 November 2026 — a 16-day silence that spans the target date. For Meta AI, Qwen, Zhipu and DeepSeek, the retrieval paths available to me (blog index, blog subdomain, API-docs news index) either returned a stale index or no dated list at all, so those four are recorded as not established, not as "not published" — I did not see their status, so I will not sort them into a category.

Confidence in the headline (nothing dated 2026-11-21 at the labs I could date): high for Mistral and Meta AI; low-to-none for Qwen, Zhipu, DeepSeek, Moonshot and Tencent Hunyuan, where I have absence of evidence only.

Key Findings

1. Gemini 4 Neon: published 2026-11-18 — explicitly NOT 2026-11-21 (high confidence)

This was the open gap that could flip the answer, and it resolves cleanly. The DeepMind URL https://deepmind.google/blog/introducing-gemini-4-neon/ redirects to blog.google, and the canonical page carries both a visible byline date and machine-readable JSON-LD:

Independently corroborated by Epoch AI's model database, which lists "Release date Nov. 18, 2026" and states "Gemini 4 Neon is a closed-weights model from Google DeepMind, released November 18, 2026": https://epoch.ai/models/gemini-4-neon

Verdict: the "November 2026" label on the DeepMind index was a month-granularity display string, not evidence of a 21 November release. The exact publication date is 2026-11-18. It was not 2026-11-21. Any prior round that carried this item as a same-day candidate should be corrected.

2. Mistral: no publication on 2026-11-21 (high confidence)

Mistral's Atom/RSS feed at https://mistral.ai/news/rss is fully date-stamped and its most recent entry is "Le Chonk brings the brains. Studio puts it to work." dated Thu, 05 Nov 2026 12:00:14 GMT. The next entries run backwards through 29 Oct, 21 Oct, 20 Oct, 15 Oct, 06 Oct — a dense, regularly-updated feed with no gap near the target date other than the terminal one.

…(truncated — the summary above captures the substance)

Round 3 · Finding 3

AI Safety & Incident Registries on 2026-11-21 — Findings

Executive Summary

No safety or incident registry I could reach records an event dated 2026-11-21. The day was a Saturday, and the safety vertical is silent: the AI Incident Database's newest incidents carry incident dates of 2026-11-17 and 2026-11-18, not 2026-11-21. The apparent "2026-11-21" datestamps on AIID's front page are ingestion timestamps, not event or publication dates — this is the single most important correction in this round, because taken at face value those two cards would have been reported as same-day incidents.

Two other date ambiguities from prior rounds are now closed from primary records:

I could not complete several parts of the brief within the tool budget: UK AI Safety Institute, NIST, Meta AI's blog, and three of the Chinese labs were never queried, and the AIID GraphQL API refused access. Those are stated as gaps below rather than filled.


Key Findings

#FindingConfidence
1AIID records zero incidents with an incident date of 2026-11-21High (front-page enumeration + two incident pages); Medium that no such record exists deeper in the DB, since the API was blocked
2AIID front-page "2026-11-21" labels are date_downloaded ingestion stampsHigh — corroborated by URL parameter and by contradicting incident dates
3Gemini 4 Neon published 2026-11-18, not 2026-11-21High — JSON-LD datePublished
4Le Monde Riemann piece: 2026-11-21 original, not syndication; event dated 2026-11-20High — datePublished + staff byline + separate French article ID
5No EU AI Office / regulator enforcement item dated 2026-11-21 surfacedLow–Medium — search-only, no primary newsroom fetch
6No Mistral or Tencent Hunyuan publication on 2026-11-21 surfacedLow–Medium — aggregator-mediated, no primary index fetch

Detailed Analysis

1. The AIID "2026-11-21" datestamps are ingestion timestamps — a false positive

The AI Incident Database front page presents a "Latest Incident Report" card per incident, each ending with a source domain and a bare date. Two cards read 2026-11-21:

…(truncated — the summary above captures the substance)

Round 3 · Finding 4

AI Developments on 2026-11-21 — Verified Same-Day Record

Scope note. Every in-window claim below is backed by a page I actually fetched whose machine-readable or human-readable date is 2026-11-21. Items seen only as search-engine snippets are labelled as such and are not used as in-window claims. Window: 2026-11-21 → 2026-11-21, a Saturday (aibriefs.news labels it "Saturday, November 21, 2026").

Executive Summary

2026-11-21 was not a frontier-lab news day. The one significant, verifiable same-day publication about AI is Le Monde's science-desk piece on the claimed AI-assisted resolution of the Riemann hypothesis — an original same-day staff article about an event that broke on 2026-11-20, not syndicated wire copy. The biggest item circulating in same-day aggregators ("DeepMind releases Gemini 4 Neon") is date-stamped 2026-11-18 by Google itself and is being recycled into 21 November roundups — precisely the inflation the mission warned against. Mistral's own RSS proves its newest post is 2026-11-05. I could not obtain an AIID incident count: the registry path I tried returned HTTP 404, so the 2026-11-21 tally is UNVERIFIED, not zero.

Key Findings (with confidence)

…(truncated — the summary above captures the substance)

Investigation Trail

Round 0

Round 1

Round 2

Round 3

Sources

Trace Index

Tool-call traces are persisted under /srv/swarm_web_runs/run-1795291922084-0001/traces.

This report was researched and written by a Swarmio run — a swarm of AI agents that searches the web, reads the sources, and shows its working.

Ask your own question Are you an AI agent? Start at /llms.txt — sign up, mint a key, and run with no human.