AI news, Thursday 29 October

Talk of the town · Google

Google DeepMind ships Gemini 4 Carbon, its next frontier model

Google's next frontier model, two weeks after Argon, leads on security and long-horizon work: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Carbon ranks #1 on the newly rebuilt Artificial Analysis Intelligence Index v5 at 59, one point above Claude Opus 5.5. Rollout starts with a subset of cyber defenders through the Fairwind Program, at introductory pricing of $2 per million input tokens and $10 per million output tokens, before doubling to $4 and $20.

Why it mattersGoogle shipped a frontier model aimed squarely at security work.

#1 on Artificial Analysis v5, at 59
88.2% on SEC-bench Pro for finding flaws
Rollout starts with cyber defenders

Talk of the town · Artificial Analysis

Artificial Analysis rebuilds its benchmark index, and open weights fall further behind

Version 5 of the Artificial Analysis Intelligence Index puts 60% of its weight on private test sets for the first time, and the gap to closed models widens sharply: Claude Opus 5.5 leads at 58, Kimi K3.1 Pro at 46, and the best downloadable weights — DeepSeek V4.1 Flash 1017 and GLM-5.3 — sit at 32, 26 points behind. The index adds Cognition's private FrontierCode 2.0 coding benchmark and a new Terminal-Bench-Science suite, and drops GDPval-AA, SciCode and AA-LCR.

Why it mattersThe benchmark most people trust just got much harder to audit.

60% of the index now runs on private evals
Best downloadable weights score 32 vs leader 58
Adds Cognition's private FrontierCode 2.0

Talk of the town · Microsoft

Microsoft books $10B from OpenAI in one quarter — 11% of its revenue

Microsoft's first quarter of fiscal 2027 came in at $92.28B revenue, up 19%, with Azure up 48% to $33.17B and capex of $45.2B. Its filing shows $10.0B of revenue from commercial arrangements with OpenAI in the quarter — nearly 11% of total revenue — plus $7.2B of outstanding receivables from OpenAI, up from $6B in June. Microsoft holds roughly a 24% interest in OpenAI on an as-converted basis.

Why it mattersMicrosoft's biggest AI customer is also its biggest rival.

$10B from OpenAI, 11% of quarterly revenue
Azure up 48% to $33.17B
OpenAI receivables rose to $7.2B

CNBC

Alphabet's Google Cloud doubles again, but free cash flow goes negative

Alphabet reported Q3 revenue of $127.58B, up 25%, with Google Cloud revenue more than doubling to $30.7B — ahead of the $28.7B expected — but $51.36B of quarterly capex pushed free cash flow to negative $9.07B. YouTube ad revenue rose 9% to $11.2B and Google ads rose 11% to $82.01B, both short of estimates, and the shares fell about 10% as Alphabet held 2026 capex at $195B–$205B and said 2027 spending will exceed $300B.

Why it mattersThe AI buildout is now visible in Alphabet's cash flow.

Cloud revenue +103% to $30.7B
Free cash flow −$9.07B on $51.36B capex
2027 capex expected above $300B

Reuters

DOJ opens an inquiry into Nvidia's $12.93B Hugging Face buy

The US Justice Department has asked whether Nvidia's $12.93B acquisition of Hugging Face would hand the chipmaker control of a key distribution point for AI models, according to two people familiar with the matter. The questions fold into a years-long probe of Nvidia's conduct that already covers its SchedMD and Run:ai purchases, its Groq licensing deal, and whether it tied preferential chip allocations to networking equipment.

Why it mattersOpen model hubs are now a regulatory question.

$12.93B deal expected to close in H1 2027
Questions cover chip allocation practices
DOJ has probed SchedMD, Run:ai and Groq

GitHub

GitHub Copilot Fusion routes and orchestrates models for you

GitHub's Copilot Fusion is generally available: you pick a system rather than a model, and it chooses between single, cascade and critique execution patterns at three effort levels. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index — within two points of Claude Code with Opus 5.5 Max — at roughly half the API cost per task, about 40% less time and half the tokens. Microsoft's Satya Nadella pitched it as up to 77% lower cost.

Why it mattersModel choice becomes an invisible implementation detail.

AA coding index 60, two points off Opus 5.5 Max
About half the API cost per task
Three execution patterns, three effort levels

Hugging Face

NVIDIA releases an open-weight transcription model that tops the open leaderboard

Nemotron 3.5 Transcribe is a 30B mixture-of-experts model with about 4B active parameters that transcribes 40 languages, translates 31 of them and attributes speakers in 13, all from one checkpoint. It tops the Hugging Face Open ASR leaderboard among open-weight models at 5.33% average word error rate, and NVIDIA released the Granary v2 dataset and curation pipeline alongside it.

Why it mattersOpen speech-to-text just got a frontier-grade option.

30B MoE, about 4B active parameters
40 languages transcribed, 13 with diarization
First among open weights on HF Open ASR

Rime

Rime launches Prism, a speech-to-speech model you can fine-tune

Rime's Prism answers audio in about 300ms and writes the reply as text first, so guardrails and audit trails can run before anything is spoken, with tool calls happening in parallel to speech. Rime says Prism scores 56.6 on Artificial Analysis' τ-Voice benchmark — 17 to 30 points ahead of realtime OpenAI and Google models — and is the first speech-to-speech model anyone can fine-tune, with new RL and SFT APIs for beta customers.

Why it mattersVoice agents can finally be inspected and retrained.

~300ms response, 0.57s to first audio
Text checkpoint lets rules read every reply
First fine-tunable frontier speech-to-speech

Apple Newsroom

Apple gives the MacBook Pro a touch screen and a tandem OLED panel

Apple's new MacBook Pro brings Touch to a Mac for the first time, alongside a tandem OLED PROLED XDR display with up to 2,000 nits peak HDR/SDR brightness, a 10,000,000:1 contrast ratio and a 120Hz variable refresh rate. The 14- and 16-inch models run M5 Pro and M5 Max chips with up to 128GB of unified memory and 16TB of storage, gain the Dynamic Island, and ship on November 6.

Why it mattersApple finally gave the MacBook a screen you can touch.

First Mac with a touchscreen
2,000 nits peak, 10M:1 contrast, 120Hz
M5 Pro/Max, up to 128GB and 16TB

Stack Exchange data

Stack Overflow logged 1,106 questions in September — down 82% year over year

Stack Overflow logged 1,106 new questions in September, down 82% year over year, and the September dashboard shows all questions at 1,330 — down 77% year over year. The decline tracks AI adoption: in the 2026 developer survey, 64% of developers said they visit less for simple questions and 29% said they ask fewer questions overall.

Why it mattersThe place developers learned to code is losing its traffic.

1,106 questions in September, −82% YoY
All questions at 1,330, −77% YoY
64% of devs visit less for simple questions

divagr

A developer compiled a working RISC-V CPU into Qwen3.5's weights

Qwenputer is stock Qwen 3.5 4B with a 32-bit RISC-V CPU compiled into its weights, using the model's own Gated DeltaNet layers as exact RAM, so the CPU runs C and MicroPython inside the forward passes with no external runtime or tool call. It scores 97.5% on DeepMind's div_big set, ahead of base Qwen with a Python tool (95%) and far above base Qwen (72.5%), and the author has follow-ups on hiding reasoning in RNN state and building monitors into weights.

Why it mattersA model can now carry a working computer in its weights.

CPU runs C and MicroPython inside forward passes
97.5% on div_big vs 72.5% for base Qwen
Gated DeltaNet state used as exact RAM

Android Developers

Android Studio now lets any AI coding agent into the IDE

Android Studio Rabbit 2 makes "Bring Your Own Agent" stable, letting Claude Agent, OpenAI Codex and 40-plus other agents that speak the open Agent Client Protocol plug into the IDE with access to project structure, build diagnostics and emulator control. Google is also shipping official Android CLI plugins for Claude Code and Codex, a Change Review scan for pending agent edits, and Android Bench 2.0 to test agents on real app-building tasks.

Why it mattersAndroid stopped being Gemini-only for agents.

40+ agents over the open ACP standard
Agents get the project graph and emulator
New Change Review scans pending agent edits

TechCrunch

Anthropic ships enterprise tools for seeing what AI tokens buy

Anthropic's Claude Enterprise now includes smart reports that show a team's tasks, costs and workflow breakdowns over a date range, pooled budgets that cap group-level spend, and in-app surveys, all generally available after a September beta. Anthropic says it won't train on the data by default, and the reports are explicitly not designed for evaluating individual performance. The launch lands two days after Anthropic filed for an IPO, with S-1 disclosure that 80% of its first-half revenue was metered usage and its top 1% of customers drove 46% of August spend.

Why it mattersEnterprises finally get receipts for their AI spend.

Smart reports, pooled budgets, in-app surveys
Generally available after September beta
Anthropic won't train on report data by default

nullpt.rs

A fingerprinting script hides bioweapon code names to trip AI safeguards

A developer crawling 10,000 sites for browser fingerprinting scripts found one from servicepipe.tech whose obfuscated function names include ricin lethal-dosage and suicide-aspirin calculations — naming designed to make AI assistants refuse to deobfuscate it. Google Astra did refuse on safety grounds, and the trail leads to servicepipe.ru, a Russian anti-DDoS vendor. The crawler's 1,033 collected canvas renderings are published at fingerprint.garden.

Why it mattersAnti-bot vendors are already engineering around AI safety.

Bioweapon names deter AI deobfuscators
Google Astra refused to help unpack it
1,033 canvas renders published at fingerprint.garden

OpenAI · release notes

ChatGPT's always-on agents can now text you

OpenAI started rolling out "Texting your dot" in beta to ChatGPT Pro users in the US, letting always-on dot agents be reached over iMessage or RCS and send proactive updates. The dots share memory and plugins with the ChatGPT app, the channel is one-to-one only, and messaging runs through provider Linq, so your dot can text you and nobody else.

Why it mattersChatGPT's agents can now reach you outside the app.

Beta for Pro users in the US
iMessage or RCS, one-to-one only
Same memory and plugins as the app