Daily AI digest, October 28

Felix Laurent 28 Oct 2026 4 min read

news

The biggest AI news of October 28 was about who gets the best models and how they reach people. Google released Gemini 4 Carbon, but only to cyber defenders for now. GitHub made Copilot Fusion generally available, so Copilot now picks the model for each prompt. Nvidia put RTX Spark laptops on sale with a new open speech model for local use. Below are the big stories plus the smaller updates from the same day.

Google gives Gemini 4 Carbon to cyber defenders first

Google announced Gemini 4 Carbon, its new frontier model for cybersecurity and long agentic coding tasks. It builds on Gemini 4 Argon, which launched two weeks earlier. We compared Argon with GPT-6 Astra in our earlier article.

Carbon is not in the Gemini app, AI Studio or the public API yet. Google is giving it first to a subset of cyber defenders and will widen access to trusted partners through its Fairwind Program. Developers, businesses and consumers come later, with no date given.

Google's own figures for the model:

#1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59
88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym for exploitation
71.0% on CWE-bench for patching
0.0% attack success on the Gray Swan prompt-injection benchmark, against 0.7% for Argon

The introductory price is $2 per million input tokens and $10 per million output tokens. It rises to $4 and $20 after the introductory period. Our full breakdown of the launch is here.

GitHub lets Copilot pick the model with Fusion

GitHub used its Universe conference to make Copilot Fusion generally available. Instead of choosing a model, the user picks Fusion and an effort level: Standard, High or Max. Fusion then sends each prompt to one model or drafts with a cheap model and escalates when needed. It can also have a second model review the first one's work.

GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index. That is within two points of Claude Code with Opus 5.5 Max, at about half the API cost and 40% less time per task. Fusion works on all Copilot plans in VS Code, JetBrains IDEs, the Copilot app and the Copilot CLI. Each model it uses is billed at its own rate, with a 30% discount on paid plans for a limited time.

GitHub also folded Copilot into GitHub Enterprise. Every licensed user gets the features and models of Copilot Enterprise, and AI usage is billed against the company's budget from the first request. Account-team customers can switch now; self-serve customers get dates in the coming weeks.

The new Octoverse report shows how fast agents spread on the platform. In the 12 months to September 30, developers made 34.3B commits, up from 12.9B a year earlier. More than 40% of pull requests in September started from an agent's draft, up from 17% in March. GitHub now counts more than 225M developers.

Nvidia puts RTX Spark laptops on sale with local models

Nvidia said RTX Spark Windows laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI are available now from $2,599. Each one pairs a Blackwell RTX GPU and a Grace CPU in one chip. The laptops get up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance. The thinnest models are 14mm thick and weigh about 3 lbs.

Nvidia shipped new open models for these machines on the same day:

Nemotron 3.5 Transcribe, an open-weight speech model for transcription, translation and audio Q&A in 40 languages. Nvidia says it ranks #1 on the Hugging Face Open ASR leaderboard
Nemotron 3.5 Bolt E4B, a model that runs in as little as 4GB of memory
FLUX 3 Image Klein from Black Forest Labs, a 5.6B-parameter image model that runs on RTX GPUs with 8GB of VRAM

DGX Spark with 64GB of memory is also now available from Nvidia's partners. We covered the transcription model in its own article.

Smaller updates

Three smaller stories followed the money and the legal fights around AI.

Meta priced its stock sale at $25.0B of common stock, up from a planned $22.5B, at $625 a share. It added $12.5B of mandatory convertible preferred stock with a 6.50% dividend. Both offerings are expected to close on October 29.
AMD will invest $1B in Meta's equity offering. AMD Helios rack systems are already installed and being tested at Meta, and Meta will deploy a custom Instinct GPU based on the MI450.
Microsoft and Epic Games keep filing takedowns against AI-made browser ports of Fortnite, Minecraft, Call of Duty and Halo, Kotaku reports. Epic CEO Tim Sweeney said the dispute "conflates a number of different things" and that Epic does not allow use of Fortnite without permission. We showed how such ports are built in our retro games piece.

Two of the three are about Meta's $37.5B raise and who is buying into it.

What the day was about

Wednesday was about control over access. Google held back its strongest model and gave it to cyber defenders before anyone else. GitHub took the model choice away from developers and let Fusion route each prompt. Its Max tier comes within two points of Claude Code at half the cost.

Nvidia went the other way and pushed models onto the user's own laptop. A $2,599 RTX Spark machine now runs a 40-language speech model and a 4GB chat model locally. On October 28 the frontier got more gated, while everyday AI moved closer to the people who use it.