AI News Roundup – 29 October 2026: Gemini 4 Carbon, OpenAI Model Spec, Mistral Large 4 and More

Here is what happened in AI today — Thursday, 29 October 2026.

Google's Gemini 4 Carbon Tops Independent Leaderboards

Google has launched Gemini 4 Carbon, a new frontier model that has moved to the top of independent evaluations. In a notable departure from the usual broad release, initial access is limited to trusted cyber defenders — a signal that Google is treating this model as a high-capability system that warrants careful rollout. The Latent Space AINews digest reports that the model leads across independent evals, making it one of the most capable publicly acknowledged models available today, even if most users cannot yet access it directly.

OpenAI Refreshes Its Model Spec and Misalignment Monitoring

OpenAI published two significant documents today. The first is an updated Model Spec that lays out five core values intended to guide more capable AI models toward sound judgement, respect for human agency, and faithfulness to human intent. The document signals OpenAI's ongoing effort to codify how its models should behave as they grow more powerful.

The second publication explains how OpenAI detects frontier model misalignment in practice. The company describes a layered monitoring approach designed to spot unauthorised agent behaviour and flag it to human responders before incidents can escalate. Together, the two releases show OpenAI leaning into transparency around safety practices at a time when agentic AI systems are becoming more capable and autonomous.

Mistral Large 4 — Nicknamed "Le Chonk" — Arrives on Hugging Face

Mistral's latest flagship model, Mistral Large 4, has landed on Hugging Face, where it has been given the affectionate community nickname Le Chonk. Few technical details are included in the Hugging Face announcement post, but the release continues Mistral's pattern of publishing capable open-weight models that developers can access and fine-tune directly. It is one of the more anticipated open model drops in recent weeks.

IBM Releases Granite Switch 4.1.1 — Model Updates as Software Updates

IBM has released Granite Switch 4.1.1 via Hugging Face, framing the release around a concept that will resonate with developers: updating AI models the same way you update software. The tagline reflects a broader industry push toward more modular, maintainable model development pipelines. While IBM has not published an extensive summary alongside the Hugging Face post, the release adds another capable open model to the ecosystem.

Apple Opens AnySpeechRecognizer and AnyDecisionModel

Apple has published two new open models on Hugging Face: AnySpeechRecognizer, aimed at speech recognition on Apple platforms, and AnyDecisionModel, focused on decision-making tasks. The announcement marks a relatively rare public open-model release from Apple and suggests the company is moving to participate more visibly in the open AI model ecosystem, particularly for on-device applications.

Project ProtoStar: Making Publicly Funded Science AI-Ready

Hugging Face has announced Project ProtoStar, an initiative aimed at making publicly funded scientific data AI-ready. The project addresses a persistent bottleneck in applying AI to scientific research: public datasets are often poorly formatted or structured for use in machine-learning pipelines. By targeting this gap, ProtoStar could meaningfully accelerate AI applications in fields that rely on government- or institution-funded research data.

Ropes & Gray Uses OpenAI for Legal Diligence

Law firm Ropes & Gray has partnered with OpenAI to develop AI tools for transactional diligence. According to OpenAI, the workflow has lawyers reviewing AI-generated grouped documents, summaries, and reports before advising clients — keeping human judgement at the centre of the process. The case study is a concrete example of how large professional services firms are beginning to embed AI into high-stakes, document-heavy work.