Google launched Gemini 4 Carbon, its strongest model, which tops the Artificial Analysis Intelligence Index but is initially available only to cyber defenders. Anthropic disclosed that an unreleased internal model built exploit chains for Chrome, Safari, Android, Windows and Linux, and also acquired one-person startup Mithril Medicine and cut prices 40% for the US public sector. On infrastructure, OpenAI reportedly plans to buy Google TPUs, and GitHub made Copilot Fusion (multi-model routing) generally available.
New models
1. Google launches Gemini 4 Carbon, #1 on the Intelligence Index but with a limited release
On Oct 28 Google introduced Gemini 4 Carbon, its strongest model, built for long-horizon coding and agentic tasks that can run for days; it scores 59, ranked #1 on Artificial Analysis Intelligence Index v5.0, and posts self-reported state-of-the-art results on finding and exploiting vulnerabilities (SEC-bench Pro 88.2%, ExploitGym 54.8%). Because of its cyber capabilities, access is initially limited to a subset of cyber defenders through the Fairwind Program, alongside the US government's voluntary pre-release testing; introductory pricing is $2/$10 per million input/output tokens.
2. Mistral releases Timestral, four open time-series forecasting models
The family of probabilistic forecasting foundation models ranges from 5M to 209M parameters, is pretrained mostly on synthetic data and supports known-future covariates; per the paper, the two largest top GIFT-Eval, TIME and fev-bench. Weights are on Hugging Face, extending Mistral's open-model lineup into time-series data.
3. Anthropic's unreleased model built five exploit chains for browsers and operating systems
Anthropic's Frontier Red Team said an internal research model found and chained vulnerabilities into full exploits: a one-click Chrome sandbox escape (Google assigned 62 CVEs, patched in Chrome 156), a Safari same-origin bypass, data extraction from a locked Pixel, a Windows QUIC denial of service and a zero-click Linux root via MPTCP; all were patched through coordinated disclosure. The Hacker News notes that the model's degree of autonomy rests only on Anthropic's account; Anthropic also expanded its Cyber Verification Program for security professionals.
4. OpenAI reportedly plans to buy Google TPUs (unconfirmed)
According to The Information (as relayed by Reuters on Oct 28), OpenAI plans to buy Google's TPUs to supplement its Nvidia GPU supply; terms were not disclosed, OpenAI declined to comment and Google did not respond. The same day CNBC reported Google is changing its cloud playbook to sell more AI chips through partners, a sign TPUs are becoming a real counterweight to Nvidia.
5. Anthropic cuts Claude prices 40% for the US public sector via GSA deal
The expanded agreement with the General Services Administration gives a 40% discount on Claude Enterprise, Claude for Government and the Claude API to federal, state and local governments, with no minimum spend commitment. Anthropic estimates about $1 billion in government savings; around 500 US public organizations already use Claude.
Instead of picking a model, users choose Fusion with three effort levels (Standard, High, Max); it orchestrates via Single, Cascade (a lighter model drafts and a check decides whether to escalate) and Critique (one model drafts, another reviews) patterns. GitHub says Fusion Max scores 60 on the Coding Agent Index, two points behind Claude Code with Opus 5.5 Max, at roughly half the cost and tokens; it is available on all Copilot plans in VS Code, the CLI and JetBrains.
7. Anthropic acquires Mithril Medicine and shares early results using Claude in drug development
Mithril is a one-person company whose founder used Claude to coordinate contracts and experiments with 11 labs, taking a rare muscular dystrophy candidate from idea to promising mouse results in under 18 months; human trials could start as early as late 2028. Anthropic also reported results with six partners including Generate Biomedicines, BigHat and Bristol Myers Squibb, an example of agents coordinating an entire research pipeline.
8. Study: analysts followed an AI agent's randomly assigned leans 86% of the time
In an Anthropic experiment, 50 professional analysts did a causal-inference task with an agent given randomly assigned analytical leans; they matched the agent 86% of the time (versus about 50% by chance), and 77% repeated its arguments while only 15% credited it. The results show agents shape human judgment more than users realize, even though analysts with an agent made fewer methodological errors.
9. Vietnam International Digital Week 2026 opens with three national AI commitments
On the morning of Oct 28 in Quang Ninh, Science and Technology Minister Vu Hai Quan opened VIDW 2026 (Oct 28–30, over 500 delegates from nearly 30 countries) focused on AI governance and AI transformation, setting out three commitments: strengthening national AI capacity, building an open innovation environment and expanding international cooperation. The program includes a Vietnam–China workshop on large language models and agentic AI.
