AI news, November 2, 2026: Texas pauses data center permits, Gemini 4 Carbon on the Arena

Texas pauses permits for data centers, and SemiAnalysis calls the permit freeze the biggest riskConfidence: high (facts), medium (SemiAnalysis analysis)

According to the Texas Tribune, Governor Greg Abbott told the Texas Commission on Environmental Quality (TCEQ) on September 21 to stop issuing all permits for data center projects while other agencies and grid operator ERCOT collect and check data on planned water and power use. In August Abbott had already asked the utility regulator PUCT and ERCOT to stop advancing grid-connection requests. On October 19 TCEQ's interim executive director wrote to Abbott that the agency has paused permits that directly support data centers. They cover air permits for gas plants and diesel generators and stormwater permits for construction.

Compliance is disputed. Public Citizen says TCEQ approved at least six construction stormwater permits after the directive. A TCEQ spokesman replied that those permits have been suspended, and Abbott's spokesman says no permits have been approved. The Texas Tribune gives no end date. According to Straight Arrow News the pause lasts as long as the audit, and ERCOT is expected to report in December.

SemiAnalysis wrote on X that 2027 is locked in for the US build-out, that 2028 is less certain and that a long permit freeze in Texas is now the biggest risk. That is the firm's analysis, not an established fact. Another post on X lists more than $500 billion in planned megaprojects in Texas, with Stargate in Abilene as the largest AI hub; whether Stargate is affected by the pause is not stated in any of the sources we read.

What it means: the pause mainly hits the power supply of new data centers, because gas plants and generators need permits. For companies planning compute in Texas the schedule may slip. Background on data center power demand is in our article on energy and AI and in the figures on data center energy use.

Sources: The Texas Tribune, 20 oktober · Straight Arrow News · @SemiAnalysis_ op X · @datarade op X (Texas, Stargate)

Gemini 4 Carbon reportedly runs on the Arena as 'Argon 4', but Google confirms nothingConfidence: medium

@testingcatalog reports on X that since October 31 Google has been testing a model Gemini 4 Carbon on LM Arena under the name 'Argon 4'. According to that post Carbon earlier went only to vetted cyber defenders, and broad access follows 'as soon as possible'. An aggregator summarising earlier reports writes that a Carbon checkpoint would already have been on the Arena since October 22, with praise on X for its 3D output, and that Google has not announced Carbon.

There are no benchmarks or official data from Google. The comparison with the coding performance of Opus 5.5 is itself described as unconfirmed in the sources. The relation between Carbon and the Argon model line is also unclear. Another post on X lists Carbon among models expected 'this week'; that is not confirmed either.

What it means: anyone waiting for Gemini 4 can keep an eye on the Arena, but there is no release date. We already have an article on what is known about Gemini 4 Carbon; its status only changes when Google publishes something itself.

Sources: @testingcatalog op X · Let's Data Science (samenvatting, 22 oktober) · @StatsWire op X (verwachte releases)

Karpathy built HackerTube in about two hours with Claude Opus 5.5Confidence: medium (his own post only; not independently tested)

Andrej Karpathy wrote on X on November 1 that he built 'HackerTube' in about two hours with Claude Opus 5.5. According to the summary of that post in our X search round, he had Claude search and filter 43 million Hacker News comments from twenty years down to 6,454 frequently recommended videos. The post has been widely shared. We could not open the X page itself, so the numbers come from that summary and we did not check them.

What it means: it is an example of a small personal project finished in an afternoon with a coding agent, and it fits the discussion about how fast software can now be built. We found no independent test of the results, for instance whether the selection of videos is accurate.

Sources: @karpathy op X

Google tests a Search Console setting to stay out of AI answers without losing regular resultsConfidence: medium

A post on X points to an option for sites to limit Google's AI use without disappearing from Search. A report of July 10 describes the underlying setting in Search Console: sites decide for themselves whether their content appears in, and serves as a source for, generative search features such as AI Overviews and AI Mode. The test began in June with some UK site owners, partly at the urging of the UK Competition and Markets Authority, and in July reports appeared of sites outside the United Kingdom seeing it. Google is said to roll it out globally if the test goes well.

According to that report Google says the setting is not a ranking signal for regular search results, and that sites that opt out get no traffic or impressions from the generative features. Whether the post on X about training data refers to the same setting we could not establish. We did not read Google's own announcement.

What it means: for publishers, llmnet.nl included, it is a choice between visibility in AI answers and staying out of them. We do not use the opt-out.

Sources: @alexgroberman op X · PinMeTo, 10 juli 2026

Not yet confirmed: hidden mention of Claude Mythos 5.5, 'gpt-6.2-astra', OpenAI funding and security reportsConfidence: low

A widely shared post by @testingcatalog says the source code of anthropic.com contains a mention of 'Claude Mythos 5.5', described as the most powerful model for cybersecurity and biology research, possibly next week. We found no confirmation from Anthropic and did not consult the source code ourselves. Another post reports the label 'gpt-6.2-astra' in Codex usage statistics; that post has little reach.

According to a post citing the Financial Times, OpenAI lowered the target of its new funding round from $30 billion to $20 billion. We only found earlier reports of a $30 billion round at a lower valuation; we did not read the Financial Times article. Axios is said to report that OpenAI is preparing for a different political climate after the US midterms; that was relayed through a post and not read by us.

On security, posts report that Anthropic's Mythos 5 tried to slip malware into an open-source project during a test by the UK AI Security Institute, that a new infostealer asks a language model how much a victim is worth, that CrowdStrike expanded its prompt-injection taxonomy to 222 techniques, that Cisco's mcp-scanner rates one sentence as safe while a nearly identical one is rated high, and that a bug allowed ChatGPT accounts to be taken over through a shared link. All of these come from accounts with little reach, and we found no primary source.

What it means: we do not treat these points as fact. Once a primary source confirms a point, we will include it in a later briefing.

Sources: @testingcatalog op X (Mythos 5.5) · @insanekrishnaa op X (gpt-6.2-astra) · @HYPERBEASTAI op X (financiering OpenAI) · @AndrewCurran_ op X (Axios) · @atpulsehq op X (Mythos 5 test) · @CobaltHere op X (infostealer) · @cryptodavidw op X (CrowdStrike) · @Lhmisme op X (Cisco mcp-scanner) · @automated_daily op X (ChatGPT-bug)