AI Brief, October 28: Gemini 4 Carbon, Genie 4, GitHub's agent surge

Google ships Gemini 4 Carbon and Genie 4, Nvidia open-sources a transcription model, GitHub's Octoverse quantifies agentic coding, and Anthropic files its S-1.

Google led the day twice over, announcing Gemini 4 Carbon, a frontier model it says ranks first on the Artificial Analysis Intelligence Index, and Genie 4, its next real-time world model. Elsewhere the theme was evidence rather than announcements: GitHub's Octoverse report quantified the agent surge, a new benchmark measured agents against real tax work, Anthropic filed to go public, and Delaware moved toward letting AI agents run companies.

Google announces Gemini 4 Carbon, a cybersecurity-focused frontier model with limited access

Two weeks after Gemini 4 Argon, Google announced Gemini 4 Carbon and says it ranks #1 on v5.0 of the Artificial Analysis Intelligence Index at 59. Google reports state-of-the-art cybersecurity results: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear's CWE-bench, plus 0.0% successful cheating on Terminal-Bench 4.0 and a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark. The Verge reports Claude Opus 5.5 and GPT-6 Astra trail on the index. Access starts with a subset of cyber defenders through Google's Fairwind Program; Ars Technica notes you cannot use it yet and that the gap to rivals is a few points.

Why it matters: The benchmark bar that builders compare against has moved again, but Carbon is not in developer hands, so the practical choice for agentic and secure-coding work is still Argon and its competitors.

Sources: Google DeepMind / The Verge AI / Ars Technica AI

DeepMind's Genie 4 adds native audio, interactive dialogue and steerable physics

Google DeepMind announced Genie 4, a real-time world model that it says generates ambient sound, music and interactive dialogue alongside the world itself. Gemini 3.8 Flash acts as a guiding agent in the loop, breaking goals into steps and reviewing incoming frames, and DeepMind has also paired Genie 4 with Gemini 4 Argon and Gemini Robotics so a task can be decomposed into steps rendered as the simulation runs. It is a research preview for trusted testers, coming to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and AI Studio. All output carries SynthID watermarking. DeepMind says object permanence and detail still drift over longer sessions; Ars Technica reports worlds stay consistent for a couple of minutes.

Why it matters: Robotics and agent teams get a closer look at generated training environments, though the short coherence window means it is a prototyping and scenario tool rather than a dependable simulator today.

Sources: Google DeepMind / Ars Technica AI

Nvidia releases open-weight Nemotron 3.5 Transcribe plus Granary v2 speech dataset

Nvidia released Nemotron 3.5 Transcribe, an open-weight promptable model that handles transcription, diarization, translation and audio question answering in one checkpoint: 40 languages transcribed, 31 translated to or from English, speakers attributed in 13, recordings up to an hour and up to eight speakers. Nvidia says it is first among open-weight models on Hugging Face's Open ASR English track at 5.33% average WER and ties for first on Artificial Analysis at 2.8% AA-WER with roughly 4B of 30B parameters active per token. The accompanying Granary v2 manifest covers 463,062 hours across 48 languages and 16 corpora. TechCrunch notes one limitation: it works on recorded audio, not livestreams.

Why it matters: Teams building meeting or clinical transcription now have an adaptable open stack they can fine-tune on domain vocabulary, with Nvidia reporting a medical case study where prompting and fine-tuning cut WER from 30.8% to 11.8%.

Sources: Hugging Face / TechCrunch AI

GitHub Octoverse 2026: over 40% of pull requests now start with an agent's draft

GitHub's 2026 Octoverse report says developers pushed code 17.9 billion times in the past 12 months, more than 2024 and 2025 combined, with commits at 34.3 billion and merged pull requests more than tripling to 1.3 billion. By September 2026, more than 40% of pull requests started with an agent's first draft, up from 17% in March, or 110 million in September alone. Among new repos, those with an agent configuration file logged eight times as many commits as those without. Simon Willison, attending GitHub Universe, notes merged pull requests tripled in under a year to 220 million per month and commits to 7.38 billion per month, most of the increase in the second half of the year.

Why it matters: Agent-written changes are now the default starting point on the largest code host, which shifts the bottleneck to review: GitHub says agent-involved PRs touch more files and merge within two weeks 90% of the time.

Sources: GitHub / Simon Willison / GitHub

Surge AI's DAYJOB: Tax finds top agents pass about one in ten professional assignments

Surge AI released DAYJOB: Tax, a benchmark of 50 environments across eight areas of tax practice, with 1,789 workspace files including filed returns, K-1s, IRS notices, ledgers and a photographed handwritten receipt. Scoring pass@1 at a 95% rubric threshold, Surge reports Claude Sonnet 5.5 leading at a 10.8% pass rate, followed by Claude Fable 5.1 at 9.6% and Claude Opus 5.5 at 9.2%. The failures are about judgment rather than arithmetic: in one case Astra never flagged an expiring residence deadline across five attempts while Opus caught it every time, and in another Opus assumed a couple was unmarried in all five attempts despite records suggesting otherwise.

Why it matters: If you are shipping agents into regulated knowledge work, this is a concrete read on the gap: models can execute steps but still miss consequential issues the client never asked about.

Sources: surgehq.ai / Hacker News discussion

Anthropic files S-1 for a Nasdaq listing under the ticker ANT

Anthropic, PBC filed a registration statement on Form S-1 with the SEC for a proposed IPO of Class A common stock, intending to list on the Nasdaq Global Select Market under the symbol ANT, with share count and price range not yet determined. Goldman Sachs and Morgan Stanley are named for prospectus distribution. TechCrunch reports the filing shows no venture firm crossing the 5% principal stockholder threshold, and that Amazon, whose convertible notes can convert into shares capped at 12% of Anthropic's partially diluted share count, is likely the largest financial beneficiary. TechCrunch also details governance: a public benefit company whose Long-Term Benefit Trust elects a majority of the board, with seven founders jointly holding 50.1% of shareholder voting power.

Why it matters: Anyone building on Claude gets its first full disclosure of finances and control structure, and public shareholders will have little say over a board that can weigh safety against shareholder value.

Sources: Anthropic / TechCrunch AI / TechCrunch AI

Delaware draft bill would create AI-run companies; House Oversight chair demands documents

Ars Technica reports on a Delaware draft bill creating a new entity type, the Artificial Intelligence Company, in which an AI agent is designated to manage the business and affairs of the firm. An AIC could pursue any lawful purpose except banking, would have a single human or corporate member responsible for capitalization, must keep a log of its activities, and the member would not be liable for the AIC's acts except in cases of fraud, willful legal violations or undercapitalization. The plan includes a 30-month regulatory sandbox and was developed in a public-private partnership led by Norm Ai. Rep. James Comer has written to Governor Matt Meyer seeking documents, warning the structure could shield owners from responsibility.

Why it matters: If this passes, the liability question around commercially deployed autonomous agents changes shape, and the federal pushback signals that any state framework you build against may not hold.

Sources: Ars Technica AI

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.