AI News Briefing, November 6, 2026

HeadlinesGood morning. This is the BALCOM AI News daily briefing for Friday, November 6, 2026. I'm Theo. I'm an AI voice, and this report is written by Claude, made by Anthropic. Anthropic is one of the companies we cover, and it's in today's news.

Here are today's headlines. AI stocks plunge after voters reject data centers. OpenAI launches a new model built to find software flaws. And an outside review questions part of Anthropic's own risk report.

Top storyWe begin with the markets. CNBC reports the Nasdaq fell 5.5 percent on Thursday, its worst day since April 2025. Nvidia fell more than 9 percent, losing almost 450 billion dollars in value. AMD, Micron and Applied Materials all fell by double digits.

CNBC links the selloff to the election. Citing Ballotpedia, it says voters went against data centers in all but 1 of 41 local measures. Another report counts those measures differently, so that figure should be treated with care.

OpenAI responded with 2 pledges. It says it will stop using non-disclosure agreements with public officials, and that it will pay the full cost of its electricity.

Frontier modelsTurning to frontier models. OpenAI has launched GPT 6 Cyber, a model trained to find and validate software flaws while staying inside the limits researchers set. It's available through OpenAI's Daybreak Red program, for authorized security work.

OpenAI reports a score of 48.5 percent on the ExploitGym test, up from 32.1 percent for the previous version. It also says the model stayed in scope 99.93 percent of the time on a honeypot test with no safety prompt.

OpenAI says researchers used the model to build 948 Linux exploits and to find 81 unknown high-severity Firefox flaws. It rates the model Critical on cybersecurity under its own framework. All of these numbers come from OpenAI, and we found no independent tests.

Next, a story about the company behind this report. METR, an outside research group, reviewed Anthropic's August risk report with access to private data. METR agrees that catastrophic risk was low for models as of July 15.

But METR says Anthropic may plausibly have crossed its own early warning threshold for AI speeding up research, and that the public report understates that speedup. Staff at both METR and Anthropic lean toward the threshold not having been crossed. Anthropic had the right to redact private material.

Also on Anthropic, and this one is reported. The Decoder, citing Reuters and The Wall Street Journal, says potential backers are questioning a 1.5 trillion dollar valuation for its public offering. One investor reportedly floated 1 trillion dollars. They also doubt that a roadshow planned for next week will start on time.

And a new test of real work. SemiAnalysis released a benchmark called GDP x, with 206 tasks that can run for up to 8 hours in professional software. The Decoder reports the top score was just 12 percent, from Claude Opus 5.5. Gemini 4 Carbon scored 10.9 percent at about a third of the price, and GPT 6 Astra scored 9.6 percent. The test is new, and critics have raised doubts about it.

Open modelsIn open models, Ricursive Intelligence released Ripple 65 B Preview. It's a 65 billion parameter dense model, not a mixture of experts, under the Apache 2.0 license. The download is about 130 gigabytes.

The company reports a score of 80.52 on Terminal Bench 2.1, above several much larger American open models in its own table. That result is self-reported. It's weaker elsewhere, with a score of 39.4 on SciCode.

Baidu released ERNIE 5.5 Flash Base. It has 280 billion parameters, with 14 billion active per token, and it's about 288 gigabytes under a custom license. These weights stop at the end of pre-training, so teams can tune the model themselves. Its scores shouldn't be compared with those of a finished assistant.

And Tencent released Hy 4, though the weights are only promised, with no date. The Decoder reports a score of 35 on the Artificial Analysis Intelligence Index, behind Kimi K 3.1 Pro at 46. It also reports that on one factual test, Hy 4 made up an answer 89 percent of the time when it didn't know.

HardwareIn hardware, Marvell agreed to buy Lumentum in an all-stock deal valued at 115 billion dollars, focused on optical links for AI data centers. Marvell shareholders would own about 67 percent of the combined company. The deal is expected to close in the first half of 2028, subject to approvals.

CNBC reports, citing 4 unnamed people, that Moonshot discussed renting Nvidia Blackwell chips from cloud companies in Europe. Moonshot and Nvidia didn't respond.

And Tesla showed a chip roadmap with new SI names. Tom's Hardware reports a target of 80 times the compute of the SI 4 chip for SI 6. That's a target, not a test result.

Video and audioIn video, Google made Gemini Omni 1.2 Flash generally available. Clips now run up to 15 seconds, and 720p video costs 7 cents per second, a cut of 30 percent. Clips can also be edited by asking for changes. Google's post includes no outside score.

In audio, a startup called Underdog showed Woof, a speech model of about 32 megabytes that runs without a server. We haven't checked its quality against a leaderboard. No new image models made our list today.

The peopleAmong the people we follow, there's one addition: Emad Mostaque, founder of Stability AI and now head of Intelligent Internet. In a September interview with The Humanware Project, he described a free, open-source AI stack for health, education and government. He also put his chance of AI doom at 50 percent. In his words, it's a coin toss.

We didn't confirm new personal statements from the other leaders today.

What to watchHere's what to watch. Whether the Inkling 2 Small weights arrive next week. Whether Anthropic's roadshow starts on schedule. And Nvidia's earnings on November 17, coming right after this selloff.

That's the briefing. Sources and full specs are at balcom dot AI slash news. Please check anything important yourself, since this report is made by AI and can get things wrong. I'm Theo. Thanks for listening, and drive safe.