Shared research report

What are the most significant developments in AI this week?

September 25, 2026

Research Report

Question: What are the most significant developments in AI this week?

Date: 2026-09-25T13:41:29.067116196+00:00

Coverage window: 2026-09-19 – 2026-09-25

Rounds: 4

Status: PARTIAL

Objective check — 4 of 5 criteria met

The run produced work, but the objective below is not fully achieved. Each unmet criterion names what is still outstanding.

Evidence: 63 claims · 49 sourced · 5 partial · 0 unsupported · 9 self-reported (no independent source) · 7 single-source

Executive Summary

As of 2026-09-25, the defining development of the week is a three-lab frontier release collision on 2026-09-21/22 that converted frontier capability into a price war: xAI shipped Grok 4.7 (Sept 21), then Anthropic shipped Claude Opus 5.5 and OpenAI shipped GPT-6 Sol and Luna on Sept 22 — Anthropic 90 minutes ahead of OpenAI — each at steep cuts to what the previous generation cost. Anthropic's Opus 5.5 carries a 40% lower run cost than Opus 5 ($4/$20 per million tokens vs. $25 output) and the best Terminal-Bench 4.0 score published this week (66.4%); OpenAI cut GPT-6 API pricing roughly 50% versus the GPT-5.6 series ($2/$10 for Sol, $0.10/$0.50 for Luna).

The week's second-order significance is that pressure arrived from outside the labs at the same time: on 2026-09-24, a coalition of roughly two dozen state attorneys general (led by New York) formally demanded federal AI regulation while insisting it not preempt state law — advocacy, not binding law — and Oracle served a force-majeure notice on Stack Infrastructure (a Blue Owl unit) over its New Mexico Project Jupiter campus, seeking to defer rent if the site misses its 2028 power target. Anthropic followed its model launch with an applied-capability result on 2026-09-23: Claude agents searched in parallel for 21 hours and surfaced a novel CRISPR-like enzyme family (ART) — a preprint whose biological function is admittedly still unknown.

No binding regulatory, enforcement, or judicial AI action dated inside 2026-09-19→25 was verified in any jurisdiction.

The week's frontier releases

DateReleaseVerified specificsSource
2026-09-21xAI Grok 4.7Release confirmed on xAI's own newsroom; machine-readable datePublished 2026-09-21T00:00:00Z. Pricing ($2/$6 per Mtok) is aggregator-level only — not primary-confirmed.x.ai/news/grok-4-7
2026-09-22Anthropic Claude Opus 5.5"First model in our new Claude 5.5 family." $4/$20 per Mtok input/output (20% less than Opus 5); cache reads $0.20 (60% less); output >30% faster; "40% less to run than Opus 5." Terminal-Bench 4.0 66.4% (vs. 52.3% Opus 5, 57.9% GPT‑6 Astra); Humanity's Last Exam 67.7% with tools; AutomationBench 40.0%. Pre-release testing by METR and Frontier Design. Anthropic states benchmark margins are a weaker guide than the scores suggest. Sonnet 5.5 and Haiku 5.5 "in the coming weeks."anthropic.com/claude-opus-5-5
2026-09-22OpenAI GPT-6 Sol and GPT-6 LunaSol ($2/$10 per Mtok) for complex/coding work; Luna ($0.10/$0.50) for high-volume tasks. ~50% cheaper than the GPT-5.6 series, and OpenAI's model pages confirm the cut is permanent, not promotional. Claimed: "about half as many mistakes as its predecessor" on OpenAI's internal factuality eval. Available in ChatGPT Work, Codex, and the API.openai.com news RSS; model page; Wayback capture
2026-09-22Anthropic status incident"Elevated errors for multiple models" — Claude Mythos 5.1, Fable 5.1 and Opus 5 affected; opened 00:57 UTC, resolved 02:35 UTC (impact stated as 00:50–02:10 UTC). Opus 5 was the last model to recover. No incidents reported Sept 23–25.status.claude.com

Policy, legal and infrastructure

DateDevelopmentWhat it actually changesSource
2026-09-24~two dozen state AGs petition Congress on AINon-binding advocacy. Led by NY AG Letitia James; demands federal safety testing standards, "uniform and transparent government-led incident response with public findings," AI safety leaders "unburdened by profit maximization," international coordination, and a prohibition on preempting state AI law plus state enforcement authority. Signatory count is disputed across sources: NY AG says 25 others, Minnesota AG says 26, ABC News says 23 states plus DC and American Samoa.ag.ny.gov; ABC News
2026-09-24Oracle force-majeure notice on Project Jupiter (New Mexico Stargate)Issuer: Oracle Corp.; recipient: Stack Infrastructure, the developer, owned by Blue Owl Capital. Seeks to defer rent if the campus misses its 2028 online target — potentially a three-year delay on rent once payments begin. It is not an event of default on the $18bn construction loan. Related debt traded below 90 cents on the dollar; Blue Owl shares fell ~3.7%. No SEC filing records it: EDGAR full-text search returns 0 hits for "Project Jupiter" and for "force majeure" at Oracle's CIK in the window.CNBC; Bloomberg via Insurance Journal; El Paso Matters
2026-09-21NVIDIA DSX ReadyQualification program for partner power and cooling products meeting NVIDIA DSX AI-factory reference designs. Launch categories: battery energy storage (Hitachi Energy, LG Energy Solution, Tesla) and cooling distribution units (LG Electronics, LiquidStack, Vertiv). NVIDIA's own caveat: "Passing qualification does not replace site-level engineering." Note: published on blogs.nvidia.com, not the press-release wire — NVIDIA issued zero press releases in the window.blogs.nvidia.com
2026-09-19Flex acquires EPC Power for $4.4BAdds 800V DC, grid-forming tech and DC-DC converters for AI data-center power; close expected Q4 2026, with a Cloud and Power Infrastructure spin-off planned for Q1 2027. Second-hand: this is a Data Center Digest summary of a Data Center Knowledge original; Flex's own release was not retrieved.datacenterdigest.net

Research, capability and platform

DateDevelopmentVerified specificsSource
2026-09-23Anthropic: Claude discovers a novel enzyme systemA new Anthropic life-sciences lab ran ~950 agents for 21 hours / 210 million tokens, gathering 200,000+ reverse transcriptases, narrowing to 3,500 candidate systems and then 20. Result: ART ("array-associated reverse transcriptases"), found mainly in bacteriophages. Anthropic states plainly: "we don't yet know its function." Preprint linked; lab works at BSL-1/BSL-2 only.anthropic.com
2026-09-24Anthropic Project Swap (economics)201 employees across six pools (SF, NYC, London…) had Claude agents barter on their behalf. Agents matched their person's book rankings on 61% of pairs; model choice mattered more to negotiating outcomes than the instructions given; the average participant would hand Claude about a third of their yearly book budget.anthropic.com/research/project-swap
2026-09-23OpenAI MentalHealthBenchListed as an OpenAI research Publication. Index card only — the item page was not read, so contents are undescribed.openai.com/news/research
2026-09-23Meta Muse Realtime AvatarDated on Meta's own research site. Headline-level TechCrunch coverage the same day reports a Muse wearable and camera-free AI glasses.research.meta.ai
2026-09-23/24Google DeepMind: three feed-dated itemsGemini 3.8 Live with Live Avatar (Sept 24); "Advancing Private AI Compute with secure, server-side memory" (Sept 23); Gemini 3.8 text-to-speech (Sept 23). Dates come from the blog RSS pubDates — DeepMind's visible blog index shows month-level dates only.deepmind.google/blog/rss.xml
2026-09-22 / 09-25Hugging Face Daily Papers highlightsSept 22: ShieldVLA (Microsoft Research, VLA safety alignment); A Lie Detector Test for Language Models; RRSI (Google, recursive self-improvement). Sept 25: Just Ask Jev (zero-shot detector of alignment failures); ExplorationBench (Tencent Hunyuan); Qwen-Planner-Agent (Tongyi-MAI). These are HF feature dates, title/affiliation level only — no abstracts or arXiv IDs confirmed.papers?date=2026-09-22; papers

Capital

DateCompanyEventTermsSource
2026-09-22Snorkel AISeries E$350M at a $3.5B valuation — nearly triple its $1.3B mark 17 months ago; co-led by Insight Partners and S32, with Addition, Lightspeed, Greylock, GV and Wells Fargo participating. Stated $375M annualized revenue run rate. Caution from the same article: peer data-labs pay 60–70% of top-line revenue to the domain experts doing the work, so headline figures are not like-for-like.techcrunch.com
2026-09-21OuraIPO (updated filing)Seeking up to $2.2B; 50M shares at $40–44; up to $14.1B market cap at the top of range. At the $42 midpoint the company nets $532.6M, of which $526.4M covers employee-share tax obligations — leaving ~$6.2M for operations. Forerunner Ventures sells its entire 9.3% stake. A wearables capital-markets event carried on AI feeds, not a verified AI-model company.techcrunch.com
2026-09-23EnvedaSeries E$311M, with Lightspeed named as a new investor on the issuer's own site. A reported ~$2B valuation was not confirmed in the issuer's text.enveda.com

Reported but not verified — treat as leads, not facts

DateClaimWhy it isn't verified
2026-09-21British Columbia attorney general sued OpenAI and Sam Altman in San Francisco federal court over the Feb 10 Tumbler Ridge school shootingSecondary aggregator text only; no court docket or AG primary record reached
~2026-09-24Australia to investigate whether an OpenAI hack of a government health website broke the lawTechCrunch headline only; body not retrieved, so neither investigating body nor statute is confirmed
2026-09-21Xiaomi MiMo-V2.6 released on Hugging Face under MIT licence; Z.ai open-sourced ZCode under Apache 2.0Aggregator text only; neither Hugging Face nor Z.ai primary record reached
~2026-09-24PrismML runs tiny LLMs on Qualcomm-powered smart glassesHeadline and relative timestamp only
2026-09-23/24Ema $77M; Bessemer $5.75B fund; Lightspeed $250M India AI fund; Morphotonics €40M; AndoHeadline and dated URL slug only — no lead investors or valuations confirmed for any
2026-09-21TSMC-anchored advanced packaging park breaks ground in KaohsiungReuters headline and URL date only; page not fetched
2026-09-22NYT: "What to Know About Recent A.I. Hacks at Google, Anthropic…"Snippet only; contents uncorroborated

Verified silence, where it matters. NVIDIA published zero press releases in the window (its newsroom's newest release is Sept 14; the window's ten items are all blog posts, and the press-release-only RSS confirms it). AMD's investor newsroom shows nothing after Aug 31. Broadcom's newest release is Sept 2; Micron's is Sept 15, with Q4 FY26 results due Sept 30. Cerebras and Groq produced nothing dateable in the window. No acquisition or partnership in the window was verified from a source other than the second-hand Flex/EPC Power item.

Analysis

The week's real story is price, not capability. All three frontier releases are cheaper per token than what they replace, and two of them shipped within 90 minutes of each other on the same day. Anthropic's own page concedes that "benchmark margins have become a less reliable guide to real-world differences" even as it publishes a table of margins — the competitive ground being contested is cost per unit of work, not the leaderboard. OpenAI's cut is confirmed permanent on its own model pages, which makes this a structural repricing rather than a promotional spike.

Non-lab pressure showed up in the same window, in two distinct forms. The AG letter is the clearest political signal: the specific demand is not just "regulate AI" but "regulate it without preempting us," which puts the coalition on a collision course with any federal framework that trades state enforcement for uniformity. Oracle's force-majeure notice is the week's sharpest reminder that AI data-center commitments carry ordinary commercial risk — the tell is that it is a rent-timing maneuver designed to preserve contractual rights, explicitly not a default, and it was disclosed through press statements rather than any SEC filing.

Capability research moved from benchmark scores to demonstrated scientific output. Anthropic's ART discovery is the most concrete applied result of the week — 950 agents, 210M tokens, 200,000 sequences scanned — and its own framing ("we don't yet know its function") is a useful calibration: the search was automated, the interpretation is not.

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.

Detailed Findings

Round 0 · Finding 1

AI Infrastructure & Safety/Capability Research — Window 2026-09-19 → 2026-09-25

Scope note and honest limitation up front: This sub-question asked specifically for (a) AI compute/chip/data-center/energy infrastructure developments and (b) AI safety/capability research papers published in the window. I was able to verify very little on both fronts from primary sources inside the window. I burned most of my search budget on queries that returned mostly pre-window or generic results, and my arXiv listing attempt 404'd (https://arxiv.org/list/cs.AI/2609 → 404, fetched 2026-09-25). I therefore report only what I actually fetched with a visible in-window date, and I explicitly flag the infrastructure and research angles as unverified / not established rather than back-filling from memory.


1. Verified in-window items (fetched pages, visible dates)

A. Daily AI news digest reporting a cluster of frontier-model releases on 2026-09-21/22

…(truncated — the summary above captures the substance)

Round 0 · Finding 2

AI Policy, Regulatory & Legal Actions, 2026-09-19 → 2026-09-25

Method / freshness note. I searched and fetched on 2026-09-25. Every in-window claim below cites a page I actually opened whose visible publication date is inside the window (2026-09-19..2026-09-25). Anything I could not verify to the window is explicitly labelled BACKGROUND or UNVERIFIED and is not counted as this week's development. This sub-question is narrow (policy/regulatory/legal only), and my search tooling degraded badly mid-run — the meta-search fallback began returning generic homepages for AI-policy queries, and direct DuckDuckGo HTML requests returned empty (HTTP 202). The in-window record I could verify is thin, and I am saying so rather than back-filling from memory.


1. Verified in-window action

1.1 United States — 26 state/territorial attorneys general formally petition Congress to enact federal AI regulation

Date: 2026-09-24. Jurisdiction: United States (state AGs). Body: Office of the New York State Attorney General, leading a multi-state coalition.

…(truncated — the summary above captures the substance)

Round 0 · Finding 3

AI Model & Product Launches, 2026-09-19 → 2026-09-25

Scope note / method: I searched the open web with date-anchored keyword queries and fetched pages directly. Two of the three items below rest on a page I actually retrieved; the third (xAI) rests only on search-result snippets and is flagged as unverified. Access to several primary sources was blocked (openai.com returned a Cloudflare interstitial, HTTP 403), and my search tooling degraded to a Bing fallback for several queries (Google DeepMind/Gemini, Meta, Microsoft, Amazon, Nvidia, Mistral, DeepSeek, Qwen, Cohere), returning generic homepages rather than news. I therefore cannot report anything for those labs, and I am not filling the gap from memory. Everything below is dated inside the window unless explicitly labelled otherwise.


1. Verified in-window model launches

1.1 Anthropic — Claude Opus 5.5 (announced 2026-09-22) — highest confidence

Source fetched: https://www.anthropic.com/claude-opus-5-5 (page states "September 22, 2026")

…(truncated — the summary above captures the substance)

Round 0 · Finding 4

AI Capital & Corporate Moves, 2026-09-19 → 2026-09-25

Research note up front (read this first). This window is unusually hard to source from the tools available in this run. The general web-search verb returned only generic homepage noise for every query I tried (OpenAI.com, gemini.google.com, chatgpt.com, etc.), so I abandoned it and worked instead from publisher category/listing pages, which do carry dated headlines. I then fetched individual articles, but three of my six article fetches and all five of my follow-up HTTP fetches timed out before the tool budget was exhausted. As a result this report distinguishes sharply between:

I did not reach Reuters, Bloomberg, The Information, or any SEC/regulatory filing in this run. Those were named in the task as targets and I could not get to them. That is a real coverage gap, not a claim that nothing happened there.


1. Executive Summary

Three items in the 2026-09-19..2026-09-25 window could be verified to Tier A, and all three are TechCrunch-sourced:

  1. Snorkel AI raised a $350M Series E at a $3.5B valuation, led by Insight Partners and S32 (Sept 22, 2026).
  2. Anthropic released Opus 5.5 (Sept 22, 2026), a frontier model release with a price cut.
  3. Oura is seeking up to $2.2B in an IPO (Sept 21, 2026) — a consumer-hardware/health deal, not a core AI-model company.

A further cluster of in-window capital and infrastructure headlines (Ema $77M, Bessemer $5.75B, Enveda $311M, Lightspeed $250M India AI fund, Oracle's New Mexico Stargate force majeure, Nexstrom 2D semiconductors, Morphotonics €40M) was visible on dated TechCrunch category pages but could not be article-verified (Tier B).

Two manifest success criteria were NOT met and I am flagging them rather than back-filling: I found only one policy/legal action (need two), and I found no acquisition or partnership deal in the window that I could verify at all.


2. Key Findings (with confidence levels)

2a. Funding rounds, IPOs and valuations (the task's focus)

…(truncated — the summary above captures the substance)

Round 1 · Finding 1

Chip & Compute-Infrastructure Announcements, 2026-09-19 → 2026-09-25

Sub-question scope: (a) in-window announcements from Nvidia, AMD, Broadcom, Micron, TSMC; (b) date-stamping and content-verification of three named infrastructure claims (datacenterdigest.net "Colossus 2 gigawatt," Oracle NM "Project Jupiter" force majeure, StorageNewsletter $20M–$80M DC-power figure).

Bottom line up front: I could not verify a single in-window chipmaker announcement from a primary source, and all three named infrastructure claims remain UNVERIFIED — two of them because the pages could not be located or read within my fetch budget. I am reporting these as explicit negatives with the exact evidence, not filling the gap from memory.


1. Chipmakers: no in-window announcement verified

AMD — verified negative from the primary record

I fetched AMD's investor-relations press-release feed directly: https://ir.amd.com/news-events/press-releases (page live, "Press Releases" index with Year filter defaulting to 2026). The newest release listed is:

Finding: AMD's IR press-release feed contains no item dated 2026-09-19 through 2026-09-25. The most recent is 2026-08-31. Caveat: this covers the IR press-release feed only; I did not reach AMD's separate corporate newsroom, so a product-level September announcement routed there would not appear in this check.

Nvidia — no in-window item found

Search run: "Nvidia announcement September 2026" (web search). Organic results were navigational only — nvidia.com homepage, Wikipedia, marketplace.nvidia.com, play.geforcenow.com, nvidianews.nvidia.com (newsroom home), Yahoo Finance quote page, and a Bloomberg topic page whose only visible date was Sep 11, 2026. No dated in-window Nvidia announcement surfaced, and I did not reach a dated Nvidia newsroom item. nvidianews.nvidia.com/news appeared in results with a "1 day ago" snippet, but I did not fetch it, so I am not treating it as a claim.

Broadcom, Micron, TSMC — searches NOT run

I must be explicit: no searches were run for Broadcom, Micron, or TSMC within this budget. Their status is therefore not "no news" — it is "not checked." Any statement that they had a quiet week would be invented.

Verdict for the success criterion: No in-window (2026-09-19→25) announcement from Nvidia, AMD, Broadcom, Micron, or TSMC could be verified. Searches run are listed in the Sources section below.


2. The three named infrastructure claims — all UNVERIFIED

…(truncated — the summary above captures the substance)

Round 1 · Finding 2

AI Safety & Capability Research, 2026-09-19 → 2026-09-25 — Non-arXiv Routes, the Opus 5.5 System Card, and the 2026-09-22 Status Incident

Scope note (read first). This answer covers only the assigned sub-question: in-window safety/capability research via non-arXiv routes, the Claude Opus 5.5 System Card, and the 2026-09-22 Anthropic status incident. I did not research the chips/compute items (Nvidia/AMD/Broadcom/Micron/TSMC), the policy/legal items (AG coalition letter, NY AG whistleblower alert, xAI Colossus 2, Oracle "Project Jupiter", StorageNewsletter DC-power figure), or Meta Muse / Alibaba chip / Mistral / DeepSeek. Those are unverified by me in this round and I make no claim about them either way — not even that they are absent.

Method note. Both web searches I ran were unproductive. search_engine_results (engine auto → Bing) for "Anthropic status page Opus 5 incident September 2026" and "Claude Opus 5.5 System Card safety evaluations" returned only generic homepage/SEO results (anthropic.com, claude.com, Wikipedia) with no usable in-window hits. Everything below therefore comes from direct primary-source fetches, which is the route that worked.


1. Anthropic — three dated, in-window research/release items (all verified by direct fetch)

…(truncated — the summary above captures the substance)

Round 1 · Finding 3

Adjudication of Three Contested Items — Primary-Document Findings

Method note and tooling caveat (read first). All three items were supposed to be settled by reading primary documents. In this run, general web search was effectively unavailable: the search_engine_results tool fell back to Bing on every call and returned only unrelated local-directory and news-homepage results (e.g. "Best Lawyers in Waterford, MI"); DuckDuckGo's HTML and Lite endpoints returned HTTP 202 (blocked) via both http_request and the browser fallback; and Mojeek was blocked with domain_blocked:www.mojeek.com after two attempts. One primary source did resolve, however: the New York Attorney General's own press-release index, fetched successfully. That gives a clean verdict on item (2) and a strong lead on item (1), but leaves item (1) numerically unadjudicated and item (3) unverified. I am reporting the negative results explicitly rather than filling them from memory.


Item 2 — NY AG AI whistleblower industry alert: RESOLVED (out-of-window)

Verdict: The alert is dated 2026-09-17, which is OUTSIDE the 2026-09-19 → 2026-09-25 window. Any report placing it in-window is factually wrong.

The Office of the New York State Attorney General press-release index lists, verbatim:

"September 17, 2026 — INDUSTRY ALERT: Attorney General James Urges Workers With Knowledge of Unsafe AI Development to File Whistleblower Complaints" — https://ag.ny.gov/press-releases (index page; release path /press-release/2026/industry-alert-attorney-general-james-urges-workers-knowledge-unsafe-ai)

The index is ordered newest-first and brackets this item between dated entries of September 18, 2026 (above) and September 16, 2026 (below), so the September 17 date is internally consistent, not a misread of position. The alert therefore falls two days before the window opens and must be treated as background, not an in-window development.

Confidence: high (primary-source index, dated, corroborated by adjacent dated entries).

Bonus in-window find from the same primary page (this one IS in-window):

"September 24, 2026 — Attorney General James Calls on Congress to Protect Americans from Unchecked AI Development" — https://ag.ny.gov/press-releases (release path /press-release/2026/attorney-general-james-calls-congress-protect-americans-unchecked-ai-development)

This September 24, 2026 release is the most likely home of the attorney-general coalition letter referenced in the manifest. Its existence and in-window date are verified from the AG's own index; see item (1) for what remains unverified.


Item 1 — AG coalition letter signatory count (23 vs. 25 vs. 26): NOT ADJUDICATED

Verdict: unresolved — I could not reach the letter or its PDF, so I cannot state a count. Do not carry any of 23/25/26 forward as settled.

…(truncated — the summary above captures the substance)

Round 1 · Finding 4

Frontier-lab & major-platform AI releases, 2026-09-19 → 2026-09-25

Scope note. This round targeted only the eight named labs/platforms, using their own primary channels (lab blogs, RSS feeds, model-release indexes, status pages). Tool budget was exhausted before I could also close the chips/compute, policy/legal and safety-research surfaces listed in the manifest — those remain open and are itemised at the end. All fetches below were made on 2026-09-25. Where a page carried no visible date, I have marked it BACKGROUND ONLY rather than dating it.


1. Verified in-window items (primary source, dated)

Meta — Muse Realtime Avatar, 2026-09-23

The only Meta AI Research item inside the window is "Bringing Your Muse to Life," dated September 23, 2026 (page byline "September 23, 2026 · 10 minute read"; JSON-LD "datePublished":"2026-09-23T00:00:00.000Z", "dateModified":"2026-09-23T00:00:00.000Z"). URL: https://research.meta.ai/blog/bringing-your-muse-to-life

Content, per that page: it introduces Muse Realtime Avatar, an audio-driven Diffusion Transformer that consumes the Muse Realtime Voice speech-token stream and generates synchronized avatar video. Stated specifics: streams 448×768 portrait video at 25 fps with ~870 ms latency (measured from end of user turn to first byte of synchronized voice+video); 8 frames per step in 20 ms on a GB200; 12 concurrent real-time sessions on a single GB200; distillation from a 40-step / 3-way-CFG teacher (120 model evaluations per chunk) to a 2-evaluation unguided student, described as a 60× reduction; compared against Runway Characters and HeyGen LiveAvatar; watermarking via Meta Video Seal; Muse app is 18+.

Important correction to the digest framing. The digest headline "Meta Muse" points at a launch that is outside the window. Meta's own research index (https://research.meta.ai/, fetched 2026-09-25) dates the Muse items as: Muse Spark 1.3 — September 2, 2026; Muse Voice Transcribe — September 1, 2026; "How We Built Safety Into Muse" — September 8, 2026; Muse Realtime Avatar — September 23, 2026. The consumer-agent launch itself is dated September 8, 2026 per Meta's own newsroom headline "Introducing Muse: The World's First Personal AI Agent Built for Everyone" (https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ — surfaced in search with a Sep 8, 2026 snippet; I did not fetch this page, so treat the Sept 8 date as snippet-level, not primary-verified). So: the Muse agent launch and Muse Spark 1.3 are BACKGROUND ONLY (out of window); the in-window Meta item is Muse Realtime Avatar, 2026-09-23.

…(truncated — the summary above captures the substance)

Round 2 · Finding 1

NVIDIA newsroom, 2026-09-19 → 2026-09-25: enumeration from the primary record

Headline finding (verified negative)

NVIDIA published ZERO press releases on nvidianews.nvidia.com between 2026-09-19 and 2026-09-25. The most recent press release in NVIDIA's own newsroom feed is dated September 14, 2026 ("NVIDIA Expands Open Source CUDA-Q Platform for Fault-Tolerant Quantum Computing"), and the newsroom archive jumps straight from the September 21–24 items (all blogs) back to September 17. No nvidianews.nvidia.com/news/... release URL falls inside the window.

What NVIDIA did publish in the window: 10 items, all blog posts

Pulled from the newsroom's own machine-readable feed https://nvidianews.nvidia.com/releases.xml (feed build stamp: Thu, 24 Sep 2026 18:08:08 GMT), which is the reverse-chronological master list. Every entry in the window is tagged Blogs; none is tagged as a press release.

#Pub date (feed pubDate, GMT)HeadlineCanonical URL
12026-09-24 14:01:10How Open Science Can Help Researchers Prepare for the Next Pandemichttps://blogs.nvidia.com/blog/open-protein-dataset/
22026-09-24 13:02:35Contain the Chaos: 'CONTROL Resonant' Launches on GeForce NOWhttps://blogs.nvidia.com/blog/geforce-now-thursday-control-resonant/
32026-09-23 15:00:51Sakeena Fiza Helps NVIDIA Hardware Succeed at Scalehttps://blogs.nvidia.com/blog/nvidia-life-sakeena-fiza/
42026-09-23 02:30:55At AI Day Singapore, NVIDIA and Partners Showcase AI Advancements Across Southeast Asiahttps://blogs.nvidia.com/blog/ai-day-singapore/
52026-09-22 12:01:47NVIDIA Isaac ROS 5.0 Advances Agentic, Open Source Robotics Developmenthttps://blogs.nvidia.com/blog/isaac-ros-5-0-agentic-open-source-robotics/
62026-09-21 18:00:19NVIDIA Launches DSX Ready to Qualify Power and Cooling Products for AI Factorieshttps://blogs.nvidia.com/blog/dsx-ready-ai-factories-power-cooling/
72026-09-21 16:42:41Why Deploying Physical AI at Scale Demands Safety at Every Layerhttps://blogs.nvidia.com/blog/physical-ai-halos-safety/
82026-09-21 16:01:49From Enablement to Execution, Egypt's AI Ecosystem Reaches Production Scalehttps://blogs.nvidia.com/blog/egypt-africa-ai-ecosystem/
92026-09-21 14:53:05AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stackhttps://blogs.nvidia.com/blog/ai-security-agent-stack/
102026-09-21 10:01:445 Companies Using NVIDIA AI for Clean Energyhttps://blogs.nvidia.com/blog/clean-energy-nvidia-ai/

Source for the list and all timestamps: https://nvidianews.nvidia.com/releases.xml (fetched 2026-09-25), corroborated item-for-item by the newsroom archive pages https://nvidianews.nvidia.com/news and https://nvidianews.nvidia.com/news?page=2.

…(truncated — the summary above captures the substance)

Round 2 · Finding 2

AI Datacenter / Compute-Infrastructure / AI-Silicon: the full datacenterdigest.net feed, 2026-09-19 → 2026-09-25

1. Executive summary

I retrieved the complete datacenterdigest.net RSS document, not the truncated portion seen before. The raw XML is 59,687 characters containing 15 <item> elements (channel <lastBuildDate> = Fri, 25 Sep 2026 09:25:03 GMT); the previous retrieval had kept only 8,192 of 29,238 parsed-text characters. Method: opened https://www.datacenterdigest.net/rss/ in a live tab and parsed the XML in-page, because plain-HTTP fetches of the feed are bot-blocked and the text extractor truncates it.

Ten of the 15 items carry a datePublished inside 2026-09-19..2026-09-25; five are dated 2026-09-18 and are excluded. The single most important structural fact: the feed contains no item dated 2026-09-22 through 2026-09-25. Its newest post is the 2026-09-21 weekly recap, even though the feed itself was rebuilt on 2026-09-25. That is a genuine content gap in this source for the last four days of the window, and it is not something I can fill from DCD.

DCD is an aggregator: each item is a digest written by Eric Lakin that links out to an original at Data Center Knowledge or Observer. The concrete figures below are DCD's summary of those originals; I did not fetch the originals, so where I cite a deal or a number I flag it as second-hand.

2. Full in-window inventory (all 15 feed items, so the split is auditable)

In window (10 items):

…(truncated — the summary above captures the substance)

Round 2 · Finding 3

AI-Semiconductor News, Broadcom / Micron / TSMC — 2026-09-19 to 2026-09-25

Scope note: This round was asked to close the chips/compute pillar for the week 2026-09-19..2026-09-25. I fetched the three vendors' primary newsrooms directly. The headline result is a negative finding that is itself verified: two of the three named vendors published no AI-silicon press release inside the window, and the week's substantive AI-silicon items sit adjacent to (not inside) their own newsrooms. I flag headline-level items as such because I did not open the underlying article — my tool budget ran out before I could.


1. Executive Summary


2. Key Findings

TSMC

…(truncated — the summary above captures the substance)

Round 2 · Finding 4

AI Compute/Silicon Findings — Window 2026-09-19 → 2026-09-25

Scope of this round: AMD's corporate newsroom (newsroom.amd.com, distinct from ir.amd.com) plus the adjacent AI-silicon slice (Intel, Google TPU, Amazon Trainium/Inferentia, Cerebras, Groq), with the DCD datacenter feed re-pulled. Every claim below is tied to a page I actually fetched; where a page was unreadable, unreachable, or undated, I say so rather than fill the gap.


1. Executive Summary


2. Key Findings

AMD — corporate newsroom (newsroom.amd.com), the flagged gap

Finding A1 — The AMD corporate newsroom carried NO Instinct/accelerator/AI-datacenter-deployment announcement between 2026-09-19 and 2026-09-25. (Confidence: HIGH, 0.85)

I fetched the AMD Newsroom homepage (https://newsroom.amd.com/, which is where https://www.amd.com/en/newsroom.html resolves) and the complete index https://newsroom.amd.com/all-news/, which states "Showing 24 of 183 articles" and lists every story in strictly reverse-chronological order from Sept 24, 2026 back to Aug 5, 2026. There is no date gap between Sept 24 and Sept 18, so the window's AMD corporate-newsroom output is exhaustively these three items:

…(truncated — the summary above captures the substance)

Round 3 · Finding 1

OpenAI's GPT-6 "Sol"/"Luna" claim — verified against non-openai.com and primary alternates

Sub-question scope: What did OpenAI announce between 2026-09-19 and 2026-09-25 regarding the models reported as GPT-6 "Sol" and "Luna"? Confirm or refute the exact model name(s), capabilities, pricing, and the claimed 2026-09-22 date.

Verdict: CONFIRMED in full. Every element of the prior round's secondary-sourced claim — the two model names, the 2026-09-22 date, the 50% price cut, and the tier roles — is corroborated by an in-window primary/alternate record (OpenAI's own news RSS, a Wayback capture of OpenAI's announcement page, and OpenAI's developer pricing pages) plus same-day Tier-1 coverage. The direct openai.com article URL still returns a Cloudflare 403, but the RSS feed fetched cleanly over plain HTTP, so the earlier blocker is resolved via alternates as instructed.


1. Exact model names — CONFIRMED

The models are GPT‑6 Sol and GPT‑6 Luna (API ids gpt-6-sol and gpt-6-luna).

No model named "Terra" was announced; 9to5Mac (Sep 22 2026 - 11:02 am PT, https://9to5mac.com/2026/09/22/openai-upgrading-chatgpt-and-codex-with-two-more-gpt-6-models/) notes "the absence of a GPT-6 Terra model."

2. Date — CONFIRMED as 2026-09-22

…(truncated — the summary above captures the substance)

Round 3 · Finding 2

Sub-question: Were xAI's Grok 4.7 release and the "Colossus 2 first gigawatt data center" claim actually published 2026-09-19→2026-09-25?

Executive Summary

Key Findings (with confidence)

…(truncated — the summary above captures the substance)

Round 3 · Finding 3

Findings

Direct answer: YES — an in-window record of the notice exists, but NOT in any of the primary records the question names. The claim is confirmed by Tier-1 and local press dated 2026-09-24, and confirmed absent from SEC EDGAR for the window. Who issued it, to whom: Oracle Corp. issued a force-majeure notice to Stack Infrastructure, the Project Jupiter developer (a portfolio company/unit of Blue Owl Capital). Reported date: 2026-09-24.

1. The notice itself — who, to whom, when (CONFIRMED, HIGH confidence)

…(truncated — the summary above captures the substance)

Round 3 · Finding 4

1. Executive Summary

This sub-question asked which AI funding/corporate and hyperscaler (AWS) announcements dated 2026-09-19 through 2026-09-25 can be confirmed from primary records. Result of this sweep:

2. 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-1790342773290-0006/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.