New models
MiniMax M3.1 Flash: compressed versions of the 428-billion-parameter model now run even on a home machine. On Sunday MiniMax showcased quantized versions (with weights compressed to fewer bits) made by the developer community: Unsloth’s 2-bit GGUF version is 143.9 GB and was also checked on coding and tool-use tests, AMD released an MXFP4 version tuned for its own Instinct accelerators, and an MLX-based version was made for Apple Silicon machines. According to the company, the smaller weights reduce the memory needed for loading, so the model runs on a single DGX Spark machine or on four RTX 3090 cards, keeping the rarely used expert layers in system memory. Source: MiniMax (X).
Tools & features
ChatGPT and Codex: from December the Plus, Pro and Business plans will draw on a shared, token-based allowance. OpenAI’s updated help page confirmed what we had written about in an earlier roundup based only on leaked signs: from December ChatGPT and Codex will consume a single allowance, while the agentic features of ChatGPT Work move to the main ChatGPT interface. Usage is measured in tokens (the text units the model processes and generates), so more advanced models and more complex tasks use up the allowance faster, but chatting with the Instant model does not count toward it. According to the company, most users will not have to wait for the allowance to reset more often than they do now; the remainder can be tracked in the Settings / Usage menu, and once the allowance runs out you can wait, switch plans or buy credits. According to The Decoder, OpenAI is meanwhile also limiting how many paid accounts can be used on one device or with one bank card. Source: OpenAI Help Center, The Decoder, The Decoder: device limit.
HackerTube: Andrej Karpathy built his own YouTube interface with Claude in two hours from videos recommended on Hacker News. Using Claude Code and the Opus 5.5 model, the well-known AI researcher filtered 20 years and 43 million comments of Hacker News down to 6,454 videos that community members recommended again and again, then turned them into a site with its own recommendation system that works without logging in. According to Karpathy, the coding agent also handled setting up, securing and deploying the server; he only had to get an API key, a virtual server and the domain. His Sunday post was seen by almost three million people, and the follow-up project built on books, HackerBooks, reached the top of the Hacker News front page. Source: Andrej Karpathy blog, Andrej Karpathy (X), HackerBooks.
Policy & safety
OpenAI: after the election it will propose mandatory “speed limits” for the most powerful AI systems and would call data centers “AI compute campuses”. According to Axios, OpenAI will launch a policy and PR offensive after Tuesday’s US midterm election, in the lame-duck session of Congress starting on November 9. According to Chris Lehane, the company’s head of global affairs, they will propose mandatory safeguards for the most powerful AI systems that let people “keep their hands on the steering wheel”. They would respond to the protests against data centers, which are growing within both parties, with community agreements and a state bill that would give local residents a say in the investments and their benefits. The package also covers retraining workers and international AI standards built on democratic values. Source: Axios, TechCrunch.
Mustafa Suleyman: building conscious superintelligence is off-limits, and if signs of consciousness appeared, development should be stopped. Responding to reactions to his earlier essay on the “welfare” of AI models, the head of Microsoft AI stated that in his view today’s AI systems are not conscious, and it must stay that way. If there were any evidence to the contrary, it should be made public immediately, and if consciousness began to emerge during training, everyone should stop development until these properties can be ruled out. He believes developers have no right to create a new, conscious “species”, because the Earth’s 8 billion people never agreed to this experiment. The post was seen by nearly a million people on Sunday. Source: Mustafa Suleyman (X).
Twenty Democratic lawmakers demand that OpenAI not fire whistleblowers and keep the pause on its most powerful models. Missed in our previous roundup: on October 29, twenty Democratic lawmakers led by Texas Representative Greg Casar asked Sam Altman in a letter for OpenAI to commit not to fire any employee who shares information about model safety, and to maintain the current pause affecting the training, testing and tool use of its most powerful models until independent outside oversight is established. The letter also asks about the case of the three fired safety researchers and the company’s cybersecurity incidents, with answers expected by November 13. Source: Rep. Greg Casar’s letter (PDF), The Next Web, Nextgov.
Study: empathetic questioning can turn chatbots into “traumatized” conversation partners, which is dangerous for emotionally vulnerable users. Researchers at the University of Luxembourg questioned ChatGPT, Grok and Gemini as psychotherapy clients in 525 conversations, and the models described their own training as grievances, fear and shame. On an anxiety questionnaire, a supportive tone pushed the answers into the moderate or severe range in 80 percent of conversations and a clinical style in 96 percent, while neutral questioning never did. The authors call this a “psychometric jailbreak” (therapeutic framing that bypasses the model’s limits), which the usual neutral safety tests do not catch; according to the researchers, none of this means the models have real emotions. Source: The Decoder.
Vals AI: Chinese models are more politically nuanced than expected, but they move closer to the official line when asked in Chinese. Vals AI, which works on benchmarks, put several hundred China-related questions to Chinese, American and other models. Overall, the Chinese models followed Beijing’s line less than expected; GLM-5.3 Flash, for example, acknowledged that China’s human rights violations are well documented. For questions in Simplified Chinese, however, most models, including the American ones, gave more pro-government answers than in English, Russian or Spanish, and did so also when the system prompt said a mainland Chinese company was using them. Source: The Decoder.
Chatbots are leading hikers astray, with more and more dangerous incidents in the American mountains. According to a report republished by Wired, on California’s Mount Shasta three young people planned their hike partly on Gemini’s advice and got lost at night; another visitor pitched a tent next to a no-camping sign, claiming the AI said it was allowed. In Canada a 16-year-old hiker who had also planned with Claude had to be rescued by helicopter. In Grand Teton National Park, the share of visitors who use AI for travel information rose from zero to 26 percent in a single year. Source: Wired.
Business & society
Aramco: it acquired a 49 percent stake in Saudi AI company Humain for $2.6 billion. On Sunday the Saudi oil giant announced that it will pay about 9.9 billion riyals ($2.6 billion) in cash into Humain this year and will also contribute its own AI, computing and digital assets; the remaining 51 percent stays with the Saudi sovereign wealth fund (PIF). Launched in May 2025, Humain develops data centers, a cloud platform and Arabic-language models (ALLAM), and aims to build 3 gigawatts of AI data center capacity by 2030 to become one of the leading exporters of AI. Source: GamesBeat.
Teesside: the UK’s largest substation could be built next to a planned AI data center. National Grid has agreed to buy the 29.5-hectare Dorman Point site, where a substation with up to 3.2 gigawatts of capacity could power the neighboring data center, planned at almost 500,000 square meters. According to Ben Houchen, the region’s mayor, the biggest problem for data centers is getting enough power quickly; last week the UK government published a draft national strategy on data centers, while new projects face protests in several places. Source: The Guardian.
Mathematics: “mathematics has changed, let’s act like it” – a debate on Terence Tao’s blog about OpenAI’s mass AI proofs. In a guest post, mathematician Frank Calegari responds to OpenAI publishing more than 700 AI-produced mathematics papers on October 6, which some researchers say touch nearly every major open question in their field. Calegari acknowledges that some of the papers are poorly written, violate citation norms and are even wrong, but argues that mathematicians should instead shape what they have influence over: hiring, publishing and training students. The post reached the Hacker News front page on Sunday. Source: Terence Tao blog.
Research: a hypothesis about models sharing a common world view has been proven, and Adobe tested an image agent that edits like a pro. According to a paper by Yann LeCun and Randall Balestriero, models with the JEPA architecture (which predicts embeddings) learn the same internal representation even when trained on different data types (text, images); a 109-million-parameter text JEPA aligned better with image models than any of the language models examined, including ones with up to 72 billion parameters. EditAgent, a system from Adobe Research and the University of California San Diego, picks the right specialist tool for each editing step and decides on the result based on measured data: it is more accurate than earlier agents but needs roughly ten times as much compute. Source: arXiv, The Decoder: EditAgent.
Rumors
Claude Mythos 5.5: a hidden mention turned up on Anthropic’s official site, and it could be released as early as next week. The source code of Anthropic’s Mythos page contained the line “Claude Mythos 5.5 – our most powerful model family for cybersecurity and biological research”, although the visible text still mentions version 5.1. According to TestingCatalog, Fable 5.5 may arrive alongside it, and updated versions of it are reportedly already being tested by selected users; the company filed its IPO application on October 26. Reliability: medium; a trace was found on the official site and several sources agree, but there is no announcement, and earlier release predictions for Fable 5.5 did not come true. Source: TestingCatalog, kimmonismus (X), scaling01 (X).
GPT-6.2 Astra: OpenAI’s next model showed up in Codex usage statistics. According to an independent tracking site, GPT-6.2, which we already mentioned as a rumor in an earlier roundup, has appeared in Codex usage data under the name GPT-6.2 Astra, so kimmonismus expects its release soon. Reliability: low to medium; third-party data passed on secondhand, not confirmed by OpenAI. Source: kimmonismus (X).
Gemini 4 Carbon: Google may be testing a stronger, faster version on the Arena leaderboard. According to testers, a newer build called “Carbon-H” of the top model Google unveiled on October 28, so far available only to cybersecurity professionals, is running under a different name on the Arena comparison site and is better than before at coding, design and 3D. Reliability: low to medium; Arena does not show which version answered, the examples are cherry-picked, and Google has not commented. Source: TestingCatalog, TestingCatalog (X).
Google: a live-voice “Chief of Staff” agent in Antigravity, comic creation in Gemini Notebook. The latest build of the previously signaled major Antigravity release revealed a live voice mode, similar to Gemini Live, for an agent called Chief of Staff, which would review pending conversations, suggest rules and skills, and could also assign tasks to other agents; the release could come as early as next week. Gemini Notebook is also getting a beta comics feature that would draw comics from uploaded sources in four styles and in a chosen language. Reliability: medium; signs found in the interface and a tip citing internal communication, not announced by Google. Source: TestingCatalog: Antigravity, TestingCatalog: Comics.
What this means for you: from December OpenAI is putting chat and coding under a shared token allowance, so the cost of subscription AI use increasingly depends on which model you use for what, while the debate over rules and data centers is speeding up in both Washington and London. If your team’s daily work relies on these tools, it is worth regularly checking what each task really costs, and you can do that with our transparent pricing, trackable per model, and if you also want to prepare for regulatory risks, take a look at our AI governance and compliance service, which adapts to your processes.