Google gave its strongest Gemini model to cyber defenders first

Google's strongest model goes to cyber defenders before customers

What happened. On 28 October Google DeepMind announced Gemini 4 Carbon, two weeks after its previous model, Argon (Kavukcuoglu, 2026). Google says Carbon ranks first, at 59, on the Artificial Analysis Intelligence Index, an independent score across many tests. It scored 88.2% on a test of finding software vulnerabilities and 71.0% on a test of fixing them. From this week a subset of cyber defenders gets access, then more trusted partners, and Google is taking part in the US government's voluntary pre-release testing. Developers, businesses and consumers get it "as soon as possible", with no date given.

The introductory price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterwards. A token is roughly a short word (Kavukcuoglu, 2026).

Why it matters. The labs now hold back their top models because of what they can do to computer systems. That makes the release schedule part of your planning: the model your competitors use may depend on which programme they were admitted to.

What to do. Ask your security provider whether it is in Google's early access group, and budget for the price doubling after the introductory period.

A Cloudflare AI service showed some customers other customers' data

What happened. Cloudflare disclosed that between 27 August and 9 September its Workers AI service returned 155 responses that contained data from other customers' requests (Flansburg & Chen, 2026). All of them answered very long prompts sent to one model, GLM 5.3 from Z.ai. The cause was a setting that left too little memory for the longest prompts, combined with a bug in SGLang, the open-source software that runs the model. A security researcher reported it on 8 September. Cloudflare took the model offline the next day and restored it with fixes. It found no sign that anyone exploited the bug on purpose.

Why it matters. When an AI provider runs one model for many customers, the shared memory is a boundary between companies, like a shared database. Cloudflare's own tests checked only that requests succeeded, not that the answers were correct.

What to do. Before you send long confidential documents to a hosted model, ask the provider how it keeps customers' data apart and how fast it discloses incidents.

Anthropic cuts the price of Claude by 40% for US governments at every level

What happened. On 27 October Anthropic expanded its agreement with the US General Services Administration, the federal buying agency (Anthropic, 2026). Government buyers get 40% off Claude Enterprise, Claude for Government and the Claude API. For the first time the discount covers state, local, tribal and territorial governments, and purchases through cloud providers will follow in the coming months. There is no minimum spend. Anthropic estimates the savings at $1 billion and says about 500 US government organisations use Claude.

Why it matters. AI vendors are competing on price for public-sector work, where contracts are long and switching is slow. European buyers can use this as a reference point in their own negotiations.

Epoch AI estimates the US received 72% of the world's AI chips since 2022

What happened. On 27 October the research group Epoch AI launched the AI Chip Imports explorer, which uses customs records from 16 countries to estimate where AI computing power ends up (Melito & Juniewicz, 2026). Of all AI computing power delivered from 2022 to June 2026, the US received 72%, China about 10% and the EU 5%. The US share of new deliveries rose to 78% in the first half of 2026. The China figure is uncertain, and chips nobody declares at customs do not appear at all.

Why it matters. Where the chips sit decides where frontier models can be trained and run, and the gap is widening. European firms that need computing power inside the EU draw on a small share of global supply.

US antitrust officials question Nvidia's purchase of Hugging Face

What happened. The US Department of Justice is asking whether Nvidia's $12.93 billion purchase of Hugging Face would give it control of a main channel for distributing AI models, according to Reuters as reported by PYMNTS (CPI, 2026). Hugging Face is the site where developers find, test and download open AI models. The questions sit within a longer review of Nvidia's acquisitions and sales practices and may lead to no action. Nvidia expects the deal to close in the first half of 2027 and says developers will not need its hardware to use the platform.

Why it matters. Many companies pick and download open models through Hugging Face. Any conditions placed on the deal will decide whether the hub stays neutral between chip makers.

The Dutch government plans 25,000 fewer civil servants by 2035, partly through AI

What happened. Dutch State Secretary Eric van der Burg told parliament he wants to cut the number of civil servants by 16%, or 25,000 people, by 2035 (Darroch, 2026). The plan relies on standard IT systems, fewer rules and tools including AI, and aims to save €3.5 billion a year. Police and defence are expected to be spared. The same letter proposes tightening the Dutch freedom of information law.

Why it matters. A European government has put a number on staff cuts linked to AI, without saying which AI uses will deliver them. Suppliers to the public sector can expect buyers to ask how much staff time their AI saves.

Sources

Anthropic. (2026, October 27). Expanding our partnership with GSA to bring Claude to all levels of government. https://www.anthropic.com/news/expanding-partnership-gsa
CPI. (2026, October 28). Nvidia's $12.9 billion Hugging Face deal draws Justice Department questions. PYMNTS. https://www.pymnts.com/cpi-posts/nvidias-12-9-billion-hugging-face-deal-draws-justice-department-questions/
Darroch, G. (2026, October 26). Van der Burg plans to use more AI so government can cut payroll. DutchNews.nl. https://www.dutchnews.nl/2026/10/van-der-burg-plans-to-use-more-ai-so-government-can-cut-payroll/
Flansburg, K., & Chen, M. (2026, October 26). Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang. Cloudflare. https://blog.cloudflare.com/kv-cache-exposure-remediation/
Kavukcuoglu, K. (2026, October 28). Gemini 4 Carbon: Our next frontier model. Google. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Melito, L., & Juniewicz, I. (2026, October 27). Introducing the AI Chip Imports explorer. Epoch AI. https://epoch.ai/latest/introducing-the-ai-chip-imports-explorer