Google unveils Gemini 4 Carbon, a cybersecurity-focused frontier model, in limited access
Google announced Gemini 4 Carbon on 2026-10-28, two weeks after Gemini 4 Argon, calling it its strongest model with state-of-the-art cybersecurity and long-horizon capabilities. It is not generally available: access starts with a small group of cyber defenders and widens to trusted partners through the Fairwind Program, with developers, businesses and consumers to follow and no date given. Google reports 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym, 71.0% on CWE-bench (patching), #4 on WebDev Arena at 1768 Elo, and a score of 59 that it says ranks first on Artificial Analysis Intelligence Index v5.0. These are Google's own figures. Introductory pricing matches Argon at $2/$10 per million input/output tokens, rising to $4/$20 afterwards. Google also reports lower cheating and prompt-injection rates than Argon. No context window is stated.
Sources & depth
Artificial Analysis releases Intelligence Index v5.0, dropping GDPval-AA
Artificial Analysis published version 5.0 of its Intelligence Index, now built from nine evaluations in four categories weighted 25% each: Knowledge Work (AA-Briefcase, AutomationBench-AA, GDP.pdf), Coding (FrontierCode 2.0, Terminal-Bench 4.0), Science (Terminal-Bench-Science 0.1, CritPt v2 Hard Subset) and General (AA-Omniscience v1.1, HLE-Diamond). GDPval-AA, a constituent through v4.3.2, was removed. Scores are not comparable with earlier versions, and Google cites v5.0 for Gemini 4 Carbon's top ranking. Nathan Lambert separately published a piece on the benchmark's design trade-offs.
Sources & depth
PyTorch 2.15 makes the rewritten NCCL backend the default and adds FP8 on Apple Silicon
PyTorch 2.15 (3,117 commits, 583 contributors) switches every nccl process group to the rewritten backend previewed in 2.14, with a flight recorder that dumps on hangs and TORCH_DIST_USE_NCCL2=0 as a rollback. It adds one-sided transports with a NIXL backend, finer FSDP2 precision control, faster NVFP4 decoding on Blackwell through NVGEMM, and FP8 plus a decode fast path on Apple Silicon that PyTorch measures at 1.32–1.34x MoE decode throughput on an M3 Ultra. CUDA 13.2 becomes the default wheel, with CUDA 13.4 wheels adding experimental Rubin support. Python 3.10 and macOS 14 support are dropped.
Sources & depth
Apple's new MacBook Pro pairs M5 Pro and M5 Max with up to 128GB of memory
Apple announced MacBook Pro models with M5 Pro (up to 18-core CPU, 20-core GPU, 64GB) and M5 Max (up to 40-core GPU, 128GB), each GPU core carrying a Neural Accelerator, plus a tandem-OLED display with touch. Apple claims up to 12x faster AI performance versus an M1 Max baseline. Pre-orders are open and the machines ship from November 6. The page gives no starting price for these configurations. The 128GB ceiling matters for running large open-weight models locally.
Sources & depth
NVIDIA's RTX Spark laptops go on sale alongside new local models
NVIDIA said RTX Spark laptops, pairing a Blackwell RTX GPU with a Grace CPU and up to 128GB of unified memory, are available now from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599. New local models include Nemotron 3.5 Transcribe (about 4.3B active parameters, 40 languages), Nemotron 3.5 Bolt in E4B and 74B-A9B sizes, Liquid AI's LFM3 family and Black Forest Labs' FLUX 3 Image Klein. A 64GB DGX Spark is also shipping.
Sources & depth