This October 29 AI briefing rounds up the five developments from the past 24 hours most worth your attention: Google launched Gemini 4 Carbon, a new frontier model built for cybersecurity; Alphabet's third-quarter cloud revenue doubled even as free cash flow turned negative; Microsoft opened GitHub's Copilot harness to enterprise developers; GitHub's annual report shows agents now initiate 40% of pull requests; and the open-weight image model MOSS-Image takes aim at the typo problem in long-text posters.
Gemini 4 Carbon launches, opening first to a small group of cyber defenders
Google DeepMind announced Gemini 4 Carbon on October 28, calling it its strongest model yet, built for cybersecurity and long-horizon tasks. The official figures: first place on version 5.0 of the Artificial Analysis Intelligence Index (59 points), 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for validating exploitability, and 71.0% on CWE-bench for patching root causes. It iterates on Gemini 4 Argon, released two weeks earlier, at the same price: $2 per million input tokens and $10 per million output tokens. The rollout is deliberately restrained: this week it opens first to a subset of cyber defenders, expanding gradually to trusted partners through the Fairwind Program, while ordinary developers and consumers will have to wait. The cybersecurity benchmarks were measured without guardrails, and the internal alignment claims still await outside verification — but handing this double-edged capability to defenders first says something in itself about how Google wants the deployment order to work.
Alphabet Q3: cloud revenue doubles, shares fall about 10% after hours
Alphabet reported third-quarter results on October 28: total revenue of $127.6 billion, up 25% year on year, a 13th consecutive quarter of double-digit growth; Google Cloud revenue of $30.7 billion, up 103%; and Search revenue of $63.3 billion, up 12%. Operating income rose 30% to $40.6 billion, but an $8.664 billion loss on equity securities dragged net income down to $26.2 billion. Capital expenditure reached $51.4 billion and free cash flow was negative $9.1 billion. The market reaction was split: cloud beat expectations while Search and operating income missed them, and the shares fell about 10% in after-hours trading. On the earnings call, CEO Sundar Pichai said the company had reached a $500 billion annual revenue run rate. The AI bill is getting heavier: revenue is rising, but cash flow is paying for compute.
Microsoft opens the Copilot harness, so agent developers need not build their own orchestration
Microsoft announced on October 28 that it is opening the Copilot harness through Microsoft Foundry — the multi-model orchestration layer built by the GitHub team that powers Copilot Code, Copilot Studio, the GitHub Copilot app, the command line, and VS Code. Developers can use it to reach more than 20 frontier models or bring their own keys, with work routed intelligently by task, plus support for MCP, agent skills, and plug-ins; the harness runtime has been ported to Rust, lowering session overhead and time to first token. Microsoft runs the agent loop and the sandbox, with governance and optimization built in. Competition is shifting from whose model is strongest to whose orchestration layer is least trouble: when models turn over every few months, a managed harness that removes the need to maintain your own routing and sandbox is a very real draw for enterprise teams.
GitHub Octoverse: developers start 40% of PRs with an agent as code volume hits records
GitHub published its Octoverse 2026 annual report on October 28. In the 12 months to September 30, 2026, developers pushed code 17.9 billion times — more than in 2024 and 2025 combined — with 34.3 billion commits and 1.3 billion merged pull requests. In September, more than 40% of pull requests were started by developers with an agent, up from 17% in March; new repositories with an agent configuration file averaged eight times as many commits as those without one; and 61% of Copilot users used more than one model in September. GitHub now serves more than 225 million developers. The eightfold figure is correlation, not proof of causation — teams that adopt agents early were probably already more active — but directing an agent first and reviewing what comes back is becoming the first move for a growing share of developers.
MOSS-Image goes open source, specialising in long text and editable layouts
The OpenMOSS team has released MOSS-Image, a 5-billion-parameter open-weight image generation model trained from scratch on a single-stream DiT architecture. It supports prompts of up to about 4K tokens, Chinese and English long-text rendering, and JSON Layout input that pins each element's position and size so a layout can be revised before generation. The technical report's figures: a mean score of 96.3 on the long-text LLTB benchmark, the highest among open-weight models; a detection-box mIoU of 0.8786 on the 7Bench layout benchmark; and a RealLayoutArena Elo of 1,698.70 for end-to-end preference, behind GPT Image 2. The weights are released under Apache-2.0, in Base and SFT versions, with support for Diffusers, Gradio, and ComfyUI. The limits are equally clear: it is a text-to-image model only, with no reference-image input and no native image editing. For posters and infographics — work with lots of text and strict placement — there is finally an open option that can be deployed locally.
