Matt Wolfe’s six-story AI news roundup moves from a proposed Anthropic IPO to model releases, an interactive world demo and a provocative mathematics claim. The useful distinction is status: GPT-6.1 Terra is available, Kimi K3.1 Pro is available through hosted products while its weights are still scheduled, Carbon and Genie 4 have limited access, and OpenAI’s ETH result is a published research claim being examined by outside experts.
The Anthropic filing is real; a $2 trillion IPO price is not set. Terra is the model to test today for repeatable computer tasks. Carbon and Genie 4 are impressive but gated. Kimi K3.1 Pro is available in apps and the API; open weights are due later. OpenAI’s ETH manuscript deserves careful review, not a “P versus NP solved” headline.
Watch Matt’s Full Roundup
Matt’s Warp Factories segment is sponsored. His benchmark comments and Genie hands-on impressions are his own tests; the figures below are checked against the linked source documents.
Anthropic Filed an S-1. The IPO Price Is Still Blank.
Anthropic announced an October 26 Form S-1 filing and an intention to list Class A shares on Nasdaq as ANT. It explicitly says the share count and price range have not been determined. “$2 trillion IPO” is a possible valuation used in outside commentary, not a number in that announcement. A listing date is not guaranteed either.
The actual SEC filing reports $16.286 billion revenue for the six months ended June 30, 2026, versus $1.163 billion a year earlier. It also reports an $80.7 billion GAAP net loss. That loss is not simply cash leaving the bank: the filing attributes $72.4 billion to fair-value changes in convertible notes, derivatives and SAFEs. Yet dismissing all cost pressure would be equally misleading. Compute and infrastructure expense was $8.945 billion for the half-year, according to the same filing.
Matt discusses an Andrew Benson X thread estimating who might gain at a $2 trillion valuation. Those per-person figures rely on assumptions about price, dilution, option exercise, vesting and individual holdings that the filing does not fully disclose. They are scenarios, not promised employee payouts. Nor would one IPO alone prove that the whole AI market is or is not a bubble.
Three Models, Three Different Access Stories
| Model | What changed | What to check |
|---|---|---|
| GPT-6.1 Terra | OpenAI positions it for visual reasoning and computer use. It is available in ChatGPT Work, Codex and the API as gpt-6.1-terra. | Standard API rates are $1 input, $0.05 cached input and $5 output per million tokens. Matt questions its cost/performance against Sol at high effort; test your own repeatable tasks instead of extrapolating from one chart. |
| Gemini 4 Carbon | Google reports an Artificial Analysis index score of 59 and stronger cyber and long-horizon results than Argon. | Initial access is for selected cyber defenders and partners, not a general launch. The announced introductory API rate is $2 input / $10 output per million tokens, rising after the introductory period. Published scores are not an independent result from this article. |
| Kimi K3.1 Pro | Moonshot offers its 3T-class model with a one-million-token context window in Kimi, Work, Code and the API, with a faster Highspeed variant. | “Open-weight” describes the planned distribution. The official blog says full weights are due by November 4, 2026; on this article’s date, use the hosted products and verify the release before planning local deployment. |
The three rows should not be read as one leaderboard. Benchmarks differ in harness, reasoning effort and task mix. Matt’s Terra discussion is a good prompt to calculate cost per accepted task, including retries and human correction, rather than only price per token. Google’s Carbon numbers and Moonshot’s Kimi results come from launch materials and should be labeled accordingly.
Genie 4: The Demo Is Real; Broad Access Is Not Yet Here
Matt gets early access to Genie 4 and navigates worlds that respond to movement and prompts. That is more than a static video clip, but it is not a dependable game engine or accurate physics simulator. Google DeepMind’s launch page includes first-party recordings and says the model is a research preview for trusted testers. It notes drift in object permanence, spatial accuracy, legible text and longer sessions. Google says access is planned for AI Ultra subscribers through Project Genie and early API/AI Studio partners; the official page is the right place to watch for that change.
OpenAI’s ETH Result: A Claim Worth Reading, Not a Shortcut to P = NP
OpenAI released a manuscript titled The Exponential Time Hypothesis is False. It claims a subexponential algorithm for a fixed-width satisfiability/counting problem that would refute ETH if the proof holds. OpenAI’s academic-collaboration announcement says an external computer-science group will work to understand and explain the result, with the AI output open for scrutiny. That is a stronger provenance trail than a headline, but it is not the same as settled community acceptance.
ETH and P versus NP are connected, but they are different statements. A subexponential algorithm can still be superpolynomial. The paper does not establish P = NP, does not immediately break modern encryption and does not hand users a practical solver for arbitrary hard problems. For the technical detail and current review status, read the manuscript and OpenAI’s collaboration note rather than a simplified “cheat code” analogy.
What Should a Reader Do This Week?
Try: Put GPT-6.1 Terra through ten low-risk, repeatable computer tasks on a test account, recording accuracy, elapsed time, API cost and corrections against the model you already use. Watch: Carbon and Genie 4 access pages rather than planning around unavailable features. Read: the SEC S-1 before repeating an employee-payout or valuation estimate. Wait for evidence: on the ETH result, separate the existence of a public manuscript from its acceptance by specialists.
Original Posts and Demos
The Genie 4 launch page embeds Google’s original world recordings; OpenAI’s Terra page contains a short launch video; and Moonshot’s Kimi technical post shows coding, 3D and finance examples. Each is a vendor demo, not an independent reproduction of Matt’s tests.
Open the original Kimi post on X. The supplied Andrew Benson link was truncated; his profile is linked above rather than embedding an unverified post ID.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Anthropic’s IPO filing and the $2 trillion scenario |
| 06:17 | Warp Factories sponsor segment |
| 07:23 | GPT-6.1 Terra |
| 12:00 | Gemini 4 Carbon |
| 13:39 | Kimi K3.1 Pro |
| 18:07 | Testing Genie 4 |
| 24:21 | OpenAI’s ETH claim |
| 28:05 | Final thoughts |
Sources and Scope
Source note, 31 October 2026: Matt’s video and the supplied transcript identify what he demonstrates and when. We checked release status against Anthropic, the SEC S-1, OpenAI Terra, Google Carbon, Google DeepMind Genie 4, Moonshot Kimi and OpenAI’s research note. We did not independently benchmark the models, reproduce Matt’s early-access Genie session, audit the S-1’s projections, or validate the ETH proof. The Warp segment is sponsored. Access, prices and research status can change after this dated snapshot.