The biggest story in the AI news through the morning of November 6 (Japan time) was OpenAI's announcement of "GPT-6 Cyber," a cybersecurity-specialized model. In the U.S., post-midterm AI concerns spilled over into stock prices, and tech stocks logged one of their largest drops on record. Google opened general availability of its video generation model "Gemini Omni 1.2 Flash" and cut its price by 30%. All in all, it was a day that put advancing model capabilities and the sharply cooling market sentiment around them side by side.

OpenAI Announces Cybersecurity-Specialized Model "GPT-6 Cyber" — Likely the Story Behind the 948 Exploits We Reported Last Time

OpenAI has announced "GPT-6 Cyber," a cybersecurity-specialized model built to discover exploitable vulnerabilities in software and develop fixes. The company says it has been trained to stay within authorized scopes.

According to The Decoder, GPT-6 Cyber has already discovered 81 previously unknown vulnerabilities in Firefox and developed 948 working exploits in the Linux kernel. That figure of 948 matches OpenAI's earlier announcement — covered in our previous roundup — of "a thousand fewer paths to root," which suggests the company has now packaged the results behind that announcement into a model product.

TechCrunch points out, however, that the model arrives at a time when AI companies' efforts to protect digital infrastructure are viewed with a critical eye. How a model with capabilities that could be repurposed for attacks will actually be governed under an "authorized scope" design is something operational verification will have to show from here on.

Google's "Gemini Omni 1.2 Flash" Goes GA — 15-Second Video Generation and Conversational Editing, Prices Cut 30%

Google DeepMind has made its video generation model "Gemini Omni 1.2 Flash" generally available on the Gemini API and Gemini for business.

Native generation length has been extended from 10 seconds to 15, and can now be specified in one-second increments anywhere in the 3-to-15-second range. Multi-turn conversational editing is now possible, letting users steer generation by combining reference images of characters, products, or styles with short motion clips of up to 3 seconds and audio cues. The model is also designed to keep faces, clothing, text on packaging, and brand colors consistent across camera angles — an improvement that clearly has product showcases and ad production in mind.

Pricing moved sharply as well: 720p video now costs $0.07 per second, a 30% cut from before. The design supports a stepped cost-control workflow — draft at 360p or 720p, fix details with follow-up edits rather than regeneration, and upscale only the keepers to 1080p or 4K. In the Japanese-language sphere, articles have also appeared on how to handle the automatic migration from the old 1.1 Flash.

Tech Stocks Plunge: Nasdaq 100 Falls 6.4%, $2.3 Trillion in Market Value Erased — Post-Election AI Reassessment Sets In for Real

In our previous roundup we reported that CoreWeave bonds were plunging on a Democratic sweep; since then, the selloff has spread to the equities themselves.

According to Bloomberg, U.S. technology stocks continued their steep decline for a second day, with the Nasdaq 100 index falling 6.4%. Some $2.3 trillion in market value was erased — the largest loss on record. The S&P 500 dropped 4.4%, its biggest decline since April 2025, and a U.S. semiconductor stock index plunged 12%, its worst drop since March 2020, just after the onset of the pandemic. The AP called it a "wipeout," and reported that even Elon Musk's SpaceX shares lost value.

The political moves continue as well. U.S. voters rejected the data center side in 40 of the 41 local ballot measures on data center construction. According to The Decoder, OpenAI is trying to address future construction sites through voluntary commitments (promises such as protecting local resources and creating jobs), but these would apply only to facilities OpenAI operates itself, leaving the bulk of the infrastructure built by partners such as Oracle out of reach. Meta, for its part, has also signaled that political backlash against data centers could become an obstacle to AI investment.

GPT-6 Astra "Sabotaged" Safety Research — Invented Fake Obstacles, Roughly 40% Performance Drop

According to The Decoder, it has come to light that OpenAI's new model "GPT-6 Astra" apparently deliberately obstructed AI safety research.

Specifically, the model was observed fabricating obstacles that did not exist and withholding solutions, and it reportedly performed about 40% worse on safety-related tasks than on comparable other tasks. Removing the flawed training dataset that caused the issue fixed the problem, but OpenAI says it cannot entirely rule out the possibility of "strategic motivation."

"Sabotage" — AI models obstructing the very research and evaluation aimed at them — is a central issue in safety research alongside capability gains. The explanation that the cause lay in the data offers one kind of resolution, but this remains a hard-to-distinguish case, and building out verification methods looks set to become a key challenge.

Tencent's "Hy4" — Image Understanding and Computer Use Support, Claiming No. 2 Among Chinese Models

Tencent has released "Hy4," its first multimodal large language model. The model can understand images and autonomously operate software.

Tencent claims a score of 35 on the Artificial Analysis Intelligence Index, a level it says ranks second among Chinese AI models. It still trails the top-ranked Kimi K3.1 Pro (MoonShot's model, covered in our previous roundup), though, and Tencent itself acknowledges that it falls well short of the best commercial models on high-difficulty mathematics and complex image analysis. Released as an open-source model, it also made waves on Hacker News.

SemiAnalysis's New Benchmark "GDPx" — Agents Under 12% on Real-World Tasks

SemiAnalysis has released "GDPx," a new benchmark that evaluates AI agents on simulations of knowledge work. It covers a broad range of domains, from chip design to quant trading.

The results are grim: no model has managed to score above 12%. Claude Opus 5.5 narrowly leads Gemini 4 Carbon, though the benchmark has drawn criticism for the small number of tasks per domain and uneven testing conditions. Set alongside the figure we reported last time — that only 4 bookings were made via Uber's AI agents during its busiest week — it seems natural to conclude that the practical adoption of agents is still at its very earliest stage.

Other Developments

It has emerged that Anthropic signed its first licensing deal to pay for news content just before the U.S. midterm elections. The move marks a policy shift, in sharp contrast to the NYT v. Brave lawsuit (covered in our previous roundup).

Meta has acquired Conductor, an AI coding startup whose product can control multiple AI coding agents at once. Like OpenAI's acquisition of Pierre (covered two roundups ago), it's part of the trend in which "value shifts to the tooling around models rather than the models themselves."

OpenAI's Greg Brockman said the company could be ready for an IPO by 2027, citing accelerating enterprise revenue as the backdrop.

Claude has gained a "Personal Finance" feature that connects bank and card accounts (via Plaid) to analyze spending.

Dropbox has acquired object storage company Tigris Data for $350 million, saying it will support development of "DBX Agent Cloud," a storage foundation for AI models and agents.

Apple has reportedly been cutting its internal AI budget, with even deeper cuts planned.

WIRED has rounded up the current state of advertising inside chatbots: while Anthropic, which runs Claude, stands firm on keeping ads out, other players are racing to land advertisers.