Agentic AI News Oct. 29, 2026

Gemini 4 Carbon. Google DeepMind shipped Gemini 4 Carbon two weeks after Argon, taking the top spot on Artificial Analysis Intelligence Index v5.0 with a score of 59 and reporting gains in coding, agentic tasks, and cyber defense. Access is initially limited to vetted cyber defenders through the U.S. government's pre-release testing process.

• Gemini 4 Carbon: Our next frontier model(opens in a new tab)
• Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity(opens in a new tab)
• Google Deepmind releases Gemini 4 Carbon just two weeks after Argon with gains across the board(opens in a new tab)
• Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use(opens in a new tab)
• Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding(opens in a new tab)

Benchmark index overhaul. Artificial Analysis' new v5 index uses 60% private test sets to reduce benchmaxxing, causing Kimi K3.1 Pro to drop from fourth to ninth and Grok 4.7 to shed 12 points. The update widens the gap between open and closed models.

• New Artificial Analysis index widens the gap between open and closed models(opens in a new tab)

Qwen4 1M context. Qwen4 models will offer a 1 million token context window, according to a qwen-code pull request. A Qwen4-27B announcement is expected at Apsara Netherlands tomorrow.

• Qwen4 1M context window(opens in a new tab)

Qwen3.8 local quants. Basalt's Strata fork recorded 5500 tok/s prompt processing and up to 275 tok/s peak generation on a single RTX 5080 with IQ3_S quantization. A new EXL3 6bpw quant of Qwen3.8-27B reaches a KLD of 0.0002 using 18.8 GB VRAM.

• [Update Benchmark] Basalt's new PRs (Strata fork) new record IQ3_S Qwen3.8 Flash Next on single RTX 5080 and 51 GB RAM with 5500 tok/s prompt processing and up to 275 tok/s peak generation (during opencode coding live session)(opens in a new tab)
• yeasah/Qwen3.8-27B-exl3 quants - lowest KLD of 0.0002(opens in a new tab)

Nvidia Nemotron 3.5 Transcribe. Nvidia released an open-weight 30B-parameter speech model with 4B active parameters that handles transcription, translation, speaker attribution, and audio Q&A for up to 40 languages. It runs on a single GPU and ships with the Granary v2 dataset and fine-tuning recipes.

• Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition(opens in a new tab)
• Voice AI in vogue: Nvidia releases new open transcription model(opens in a new tab)
• NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt(opens in a new tab)
• nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B · Hugging Face(opens in a new tab)

Hardware sticker shock. RTX 5080 cards are now selling for nearly $2,000, while a Mac Studio M5 Ultra 512GB starts at $22,299. Enthusiasts are converting DDR5 SODIMM kits to desktop DIMMs to fight RAM costs, and a decommissioned 9,600-GPU A100 cluster is listed on eBay for $60 million.

• NVIDIA RTX 5080 is now selling for nearly $2,000 at US retailers, $800 higher in just two months(opens in a new tab)
• Mac Studio M5 Ultra 512gb Starts at $22,299.00 USD(opens in a new tab)
• Electronics hobbyist converts $120 DDR5 SODIMM kit into ‘$1000’ desktop kit to fight RAM crisis — ICs from SK hynix DDR5-5600 16GB modules used to create DDR5-7800 desktop sticks(opens in a new tab)
• IT vendor lists massive $60 million AI supercomputer on eBay with 9,600 Nvidia A100 GPUs and 4,200 AMD EPYC CPUs, plus free shipping — 300 racks of decommissioned GPUs on offer(opens in a new tab)

Local LLM eval. A 50-prompt benchmark had four frontier models grade 39 local LLMs, producing rankings for the most capable model and the best foundation model to fine-tune.

• 5 Frontier LLM's grade 39 local LLM's on 50 prompts(opens in a new tab)

Agent memory and scheduling. LangChain published guides for evidence-grounded agent memory using compiled wikis and graphs, reporting 27% higher accuracy and 58% lower cost per correct answer. It also showed how cron-based schedules power proactive, always-on agents in Managed Deep Agents.

• How to build evidence-grounded agent memory with OpenWiki and Graphs(opens in a new tab)
• How to Build Proactive, Always-On Agents with the Schedules SDK in Managed Deep Agents(opens in a new tab)

Supabase for agents. Latent Space interviewed Supabase's founders about why frontier coding agents default to Supabase as an app backend, citing local backend without Docker, Git-tracked schemas, and compute next to the database as key friction removers.

• Why Agents Love Supabase(opens in a new tab)

Google agent integrations. Google launched a Connected Apps Platform beta to let third-party apps integrate with Search and Gemini. Android Studio's 'Bring Your Own Agent' supports 40+ external coding agents, and a new Gemini Nano 4 Agentic API enables on-device agent functions.

• As AI agents reshape the web, Google opens Gemini and Search to more apps(opens in a new tab)
• Google opens Android development to third-party AI agents and unveils new Gemini APIs(opens in a new tab)

MiniMax MCode Stack. MiniMax open-sourced MCode Stack, a set of 15 agent skills and checklists its MiniMax Code team uses daily. The MIT-licensed skills include long-command audits and run in MiniMax Code or Claude Code.

• MiniMax Open-Sources MCode Stack: 15 Agent Skills Its Coding Team Uses Daily(opens in a new tab)

OpenAI-ServiceNow expansion. ServiceNow will bring OpenAI models and agents into its workflows and connect ServiceNow actions to ChatGPT and Codex. IT teams can investigate incidents, assess change risk, and update records without leaving ChatGPT.

• OpenAI and ServiceNow expand partnership across enterprise workflows(opens in a new tab)

Microsoft's AI revenue. Microsoft disclosed Azure revenue of $33.2 billion, up 48%, and $10 billion in OpenAI-related revenue for the quarter. Total revenue reached $92.3 billion, marking ten straight record quarters.

• Microsoft just shared Azure’s actual revenue in dollars: $33.2 billion(opens in a new tab)
• Microsoft recorded $10 billion in revenue from OpenAI(opens in a new tab)
• Microsoft hits 10 quarters of record-breaking revenue in a row(opens in a new tab)

Google AI search ads. Alphabet reported $127.6 billion in quarterly revenue, but Search ad growth slowed to 12%. Google is expanding ads inside AI Overviews and AI Mode to international markets to build a monetization footprint.

• Google expands AI Search ads globally as core ad growth slows(opens in a new tab)

Anthropic spend analytics. Anthropic launched reporting tools that let enterprises see what their token spend buys, with team-level analytics, budgets, and ROI recommendations. Data is not used for training by default.

• Anthropic just launched tools to show companies what their AI spending is buying(opens in a new tab)

Claude for Legal. Anthropic added pre-built templates and automatic source verification to Claude for Legal, linking statements to original sources before lawyer review. The platform now has over 60 legal integrations.

• Claude for Legal now verifies sources automatically, then hands the file to a lawyer(opens in a new tab)

GitHub Octoverse. GitHub's Octoverse 2026 shows merged pull requests tripled to 220 million per month and commits reached 7.38 billion per month, reflecting the impact of agentic engineering.

• GitHub Octoverse 2026(opens in a new tab)

Meta terms backlash. Meta likely delays its new business terms past November 20 after publishers raised AI-training concerns. The company says it will not change policy without full consultation with industry groups.

• Meta says it’ll probably delay its controversial new business terms again(opens in a new tab)

xAI home buyouts. An xAI subsidiary is offering to buy hundreds of homes near its Mississippi data center in exchange for residents not joining lawsuits over noise.

• xAI is buying out homeowners who promise not to sue over data center noise(opens in a new tab)

AI backlash grows. Instagram's Adam Mosseri says negativity around AI will get worse, driven by fear of being left behind. Marissa Mayer's AI app demo drew criticism for being out of touch with everyday users.

• Adam Mosseri says negativity around AI is ‘going to get worse’(opens in a new tab)
• AI execs may be too out of touch to build for the general public(opens in a new tab)

Enterprise AI reality. An MIT Technology Review poll finds only 16% of organizations see transformative AI value, 12% have agents in production, and 38% are still piloting use cases.

• Creating AI value with enterprise context(opens in a new tab)

AI security escalation. Wiz's CTO says AI is accelerating vulnerability discovery and exploitation, as shown by PHP2Shell, and manual patching won't scale. South Korean brokerages and Bithumb were hit by fresh cyberattacks, with an earlier wave linked to AI pen-testing tool Artex.

• Wiz’s CTO on PHP2Shell and the coming ‘vulnpocalypse’(opens in a new tab)
• The ‘Vulnpocalypse’ is coming for the internet’s aging code, says Wiz’s CTO(opens in a new tab)
• South Korea’s financial industry hit by fresh wave of cyberattacks(opens in a new tab)

Local model roots TV. A user ran Qwen 3.8 Flash Next on an underspecced PC through the opencode harness to root an LG TV and remove spyware and adware.

• Used Qwen 3.8 flash next on underspeced PC to Root and gut LG TV of spy and adware(opens in a new tab)

Doom inside Qwen. A project got a basic version of Doom running on the virtual machine created inside Qwen model weights, inspired by the qwenputer project.

• I ran Doom inside the Qwen model weights!(opens in a new tab)

Single-neuron identity. An experiment tweaking one neuron at a time in Qwen3-8B and 14B found about 4% of interventions caused the model to claim a different identity, raising skepticism about identity-based distillation claims.

• Single-Neuron Identity Shifts(opens in a new tab)

Vibe-coded games. Simon Willison argues October 2026 is an inflection point for vibe-coded games, where Claude Opus 5.5 can create genuinely fun gameplay loops without the user reading code.

• The October inflection point for vibe game development(opens in a new tab)