Frontier AI, week of 30 October 2026
Gemini 4 Carbon, GPT-6.1 Terra, Anthropic's IPO filing, OpenAI's new Model Spec and Mistral's trillion-parameter open weights
A week where the labs talked as much about security and rules as about raw capability. Google led with a cyber-defence model, OpenAI rewrote how its models should behave, Anthropic filed to go public, and the open-weight world got a trillion-parameter newcomer. Here is what happened, grouped by theme, with sources.
Models and releases
Google announces Gemini 4 Carbon, cyber defenders first
On 28 October Google announced Gemini 4 Carbon, two weeks after Gemini 4 Argon. Google says it ranks first on version 5.0 of the Artificial Analysis Intelligence Index (score 59) and reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, plus 0.0% on the Gray Swan prompt-injection benchmark. Access starts with a subset of cyber defenders through Google’s Fairwind Program, and Google says it is taking part in the US government’s voluntary pre-release testing. It will launch at $2 per million input tokens and $10 per million output, rising to $4 and $20 after the introductory period.
The truth: The headline isn’t the leaderboard, it’s the rollout: a model good enough at exploitation that Google won’t hand it to everyone yet. That’s a sensible call, and a sign of where frontier capability now bites.
Source: Google: Gemini 4 Carbon, our next frontier model
OpenAI releases GPT-6.1 Terra for cheaper computer use
On 26 October OpenAI released GPT-6.1 Terra in ChatGPT Work, Codex and the API, a mid-sized model aimed at visual reasoning, computer use and writing. OpenAI says its API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, at $1 per million input tokens and $5 per million output, and that its error rate on an internal set of hard prompts fell from 53.6% to 16.7% at low reasoning effort. OpenAI also says Terra meets the “Critical” cybersecurity threshold in its Preparedness Framework.
The truth: Agents that click through forms and invoices all day live or die on cost per task, and that is what OpenAI is selling here. The comparisons with rival models are OpenAI’s own numbers; wait for independent runs before believing the “9x cheaper” maths.
Source: OpenAI: Introducing GPT-6.1 Terra
Research
DeepMind’s Genie 4 simulates worlds, with sound
Google DeepMind announced Genie 4 on 27 October, a real-time interactive world model you can steer with text prompts. New this time is native audio and dialogue, with Gemini 3.8 Flash acting as a guiding agent in the loop, and output carries SynthID watermarks. It is a research preview for trusted testers, coming next to AI Ultra subscribers in Project Genie and to early-access API users; DeepMind lists drift and loss of object permanence over longer sessions among its limitations.
The truth: The flashy part is the photorealism. The useful part is cheap, endless practice worlds for robots and agents, and that only works if the physics holds up, which DeepMind says it doesn’t fully yet.
Source: Google DeepMind: Genie 4, a new generation of world simulation
Anthropic reports early results using Claude in drug development
On 27 October Anthropic described partnerships with six organisations developing medicines. Examples it gives: Generate Biomedicines tested a Claude-proposed CAR-T component about 70% shorter that let immune cells make roughly 10 times as many receptors in lab tests; Divergence Labs got about 35,000 antibody designs in a month; Enveda says manufacturing-document review time fell 89%. Anthropic also bought Mithril Medicine, a one-person muscular dystrophy company, and says human trials could start as soon as late 2028.
The truth: Anthropic itself says this is “mostly preclinical”, so no patient has been helped yet. The more interesting claim is that the slow parts of medicine are paperwork and coordination, and that is exactly where these models already work.
Source: Anthropic: Early progress on accelerating drug development with Claude
Agents and tools
Gemini Robotics ER 2 reaches Google Cloud in preview
On 27 October Google Cloud put Gemini Robotics Embodied Reasoning 2 into public preview, a model meant to act as a robot’s high-level “brain” that plans tasks and calls lower-level control models as tools. Google says it pinpoints key moments in video, such as when to stop pouring, with 91.3% accuracy, and can read more than ten kinds of instruments like dials and thermometers. Boston Dynamics and A*STAR are among the quoted early users.
The truth: Agents are leaving the browser. A robot that can tell whether a step actually worked before moving on is a bigger deal than one that can do a flashy demo once.
Source: Google Cloud: Gemini Robotics ER 2 is now available on Google Cloud
Policy and safety
OpenAI rebuilds its Model Spec around five values
On 29 October OpenAI updated its Model Spec, the public document describing how its models should behave, reorganising it around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. It adds guidance on context, oversight and agency, retires most of the detailed style rules, and now calls the model “GPT” rather than “the assistant”. OpenAI says its models “do not yet fully reflect” the new Spec.
The truth: Read that last line twice: this is a target, not a description. Still, publishing the rules in plain language is the right move, because it gives everyone something concrete to hold the company to.
Source: OpenAI: An updated Model Spec for more capable and beneficial AI
Anthropic tells California courts should be able to block risky models
Reuters reported on 28 October that Anthropic told Governor Gavin Newsom, in a 15 October letter, that officials should be able to ask a court to block the release of a model that poses a “serious risk”. The letter says whole-model “kill switches” could give a “false sense of reassurance”, and favours targeted steps like disabling one agent or cutting off one customer. OpenAI sent its own recommendations on 20 October.
The truth: Notice who gets the power in Anthropic’s version: judges, case by case, rather than a blanket shutdown rule. Every lab now agrees oversight is coming; the fight is over who holds the off switch.
China’s draft AI export curbs: sell access, keep the models
China published draft AI export rules on 16 October, just before our window; Reuters analysed the fallout on 29 October. The draft adds 23 technologies to the restricted export list, while overseas users can still access Chinese AI services. Models are covered if training passed 10^26 calculations, they have at least 2 trillion parameters, or they rank in the global top 10. Comments close on 15 November, and a source told Reuters adoption is expected by year end.
The truth: It mirrors US chip controls in reverse: Washington guards the hardware going in, Beijing guards the models going out. The “top 10” test is the part to watch, because it ties the rules to a moving leaderboard rather than a fixed size.
Source: Reuters via The Economic Times: China’s new export curbs draw a line around its AI crown jewels
Industry and compute
Anthropic files to go public
On 26 October Anthropic filed a Form S-1 with the US Securities and Exchange Commission for a proposed IPO, planning to list on the Nasdaq Global Select Market under the ticker “ANT”. The number of shares and the price range have not been set, and Goldman Sachs and Morgan Stanley are named for the prospectus.
The truth: The filing matters more than the ticker. A public company has to disclose spending, risks and losses in detail, so expect the S-1 to say more about the real economics of frontier AI than any launch post.
Source: Anthropic: Anthropic files S-1 with the SEC for proposed initial public offering
Data centres face voters in ten US states
Reuters reported on 29 October that voters in 10 US states will decide on data-centre restrictions in the midterms, with at least 41 measures across 38 local jurisdictions. Ohio has the most, 21 proposals; one in Sunbury, banning data centres over 25 megawatts, could derail a proposed $2 billion Amazon project. Cushman & Wakefield told Reuters US capacity under construction still rose 48% between December 2025 and June 2026.
The truth: The limit on AI may end up being a town council, not a chip. Local votes won’t stop the build-out on their own, but they make power and land political questions, not just engineering ones.
Source: Reuters via Investing.com: AI data center boom faces backlash at US ballot box
Open source and open weights
Mistral releases weights for Mistral Large 4
On 29 October Mistral released the weights of an improved Mistral Large 4, a multimodal model with about 1 trillion parameters, trained from scratch in Europe. Mistral reports 48% on the Artificial Analysis Coding Agent Index (tied with GLM-5.3 while using 47% fewer tokens), 41% on Terminal-Bench 4.0 and 51% on the Artificial Analysis Cyber Index, with a 96.6% refusal rate on malicious cyber requests. The licence is a modified MIT licence under which companies offering it as a hosted model service need a separate commercial licence; API pricing is $0.70 per million input tokens and $2.10 per million output.
The truth: A frontier-scale model you can host yourself is a genuine sovereignty story for Europe. But Mistral also admits it saw “potential for uplift to conduct real attacks” and is tightening guardrails, and guardrails on downloadable weights are only as strong as the person running them.
Source: Mistral: Mistral Large 4 open weights
Liquid AI’s LFM3: six small open models on one backbone
On 27 October Liquid AI released LFM3, six open-weight models of about 2 billion parameters for text, vision, audio, combined “omni” input, and fast structured decisions, all built from one base model. Liquid says the text model fits in about 1 GB at 4-bit precision, has a 256K-token context, and can be fine-tuned on a single GPU; it reports a WebArena-Verified score of 38.42 against 3.02 for its previous 2.6B model.
The truth: Not everything needs a trillion parameters. Small models you can tune on one card and run on a phone or robot are where a lot of practical AI will actually live.
Source: Hugging Face: Build specialized multimodal models with LFM3
Sources were read on 30 October 2026. Spotted an error? That’s the point: tell me.