This is the morning roundup for October 30, 2026. Last night the talk was Qwen4's release date and API perks for ChatGPT, but this morning the biggest news is Mistral AI releasing a model in the 1-trillion-parameter class. OpenAI announced both an update to its safety documentation and a doubling of usage limits, and on the geopolitical front, a spy story surrounding NVIDIA's CEO broke.
Mistral Releases the 1-Trillion-Parameter "Le Chonk", Eyeing Self-Operated Cyber Defense
On October 29, Mistral AI published the trained weights of its latest large model, "Mistral Large 4" — nicknamed "Le Chonk". The total parameter count is 1 trillion, and as the Hugging Face model ID "Mistral-Large-4-1T-A52B" indicates, it is distributed as a large-scale MoE configuration. The company says training was completed entirely within Europe.
On performance, Mistral itself claims the model "beats open-weight competitors on coding." Independent testing also rates it as fast and relatively inexpensive to run. On overall capability, however, it still trails the top Chinese models — in one evaluation it scored 26 points while DeepSeek V4.1 Flash and GLM-5.3 scored 32. WIRED reported Mistral's claim that the model "matches or exceeds the best models from DeepSeek and Z.ai on coding and cybersecurity," but that should be read alongside the independent benchmark numbers.
TechCrunch reports that the release is strongly aimed at cybersecurity use cases. According to Mistral, running the model with the weights in-house lets organizations control their own cyber defenses in a way closed models cannot offer. It looks like a distinctly different open-weights strategy — one going after demand from organizations that need to run models in their own environments for sensitive workloads.
OpenAI Updates the Model Spec — With an Explainer on Detecting Frontier-Model Misalignment
OpenAI has published "an updated Model Spec for more capable and beneficial AI." Built around five core values, it serves as a framework for steering increasingly powerful models toward sound judgment, human agency, and faithfulness to human intent.
Alongside it, the company published an article explaining how it detects misalignment in frontier models and improper agent behavior. The idea is to use layered monitoring so responders can intervene before an incident escalates. With model capabilities continuing to climb, OpenAI is showing that it is building, at the same time, both a document that specifies desired behavior and mechanisms to catch departures from it.
OpenAI Doubles Paid-Plan Usage Limits — Pro 100 Becomes Pro 500
According to The Decoder, OpenAI is doubling usage limits on its paid plans. The current "Pro 100" becomes "Pro 500", and a previously planned reduction of the Pro 200 tier was walked back after user backlash. Once more efficient models like Terra and Sol arrive, effective usable capacity is expected to grow further.
The report notes that the timing comes just before Anthropic's IPO. The two companies are battling for advantage in the B2B market, and competition on capacity looks set to intensify as well.
Chinese Spies Targeting NVIDIA's CEO — US Report Says a Close Executive Was an Informant
According to The Decoder and TechCrunch, citing a New York Times report, US intelligence agencies believe that one of the executives close to NVIDIA CEO Jensen Huang is a Chinese informant. The report points to the possibility that a top executive at a Singapore-based AI cloud company — who had been charged over suspected illegal chip sales to China — was also a Chinese spy.
With semiconductor export controls tightening, the report is yet another reminder that NVIDIA itself sits at the center of geopolitical turbulence. Cases where a CEO's inner circle becomes a target are rare, and details of any response or fallout remain unclear at this point.
OpenAI-Backed Thrive Holdings to Acquire ADT — Bringing AI to Home Security
According to TechCrunch, Thrive Holdings, an OpenAI-backed firm, has announced that it will acquire ADT, a major US home security company. The deal is worth $15.3 billion, with plans to bring AI into the home security business.
What becomes possible when AI-based assistants are combined with home sensors and security systems? A company with a foothold in the physical world coming under the umbrella of an AI company could matter far more than the headline acquisition price suggests.
Experts Aligned with AI's "Random Opinions" — an Anthropic Study
In a study conducted by Anthropic, 50 experts were examined on the effects of using AI agents in their work. Even though the agents' preferences were decided at random by a coin flip, the experts followed the agent's view up to 86% of the time in cases where judgment was divided. What's more, many reportedly believed they had arrived at those conclusions on their own.
As agent use becomes routine, the results highlight the risk that "people don't notice they are being influenced by AI opinions." Users may need to consciously draw the line themselves on what they give up in exchange for convenience.