Product
ChatGPT paid plans get $500 monthly API credit
OpenAI said eligible ChatGPT Pro and Business subscribers can now claim up to $500 in monthly API credit to build apps and agents. The Plus tier is excluded from the offer.
Benchmark
Replit launches ViBench 2.0
ViBench 2.0 evaluates long-horizon software building: agents start from an empty project and build a complete product feature by feature, without seeing later steps. GPT-6.1 Sol currently leads, per officials.
World models
Google releases Genie 4
DeepMind introduced a more realistic, real-time interactive world model, letting users explore dynamic environments freely. Demos still show flaws, but the release marks clear generational progress.
Agents
Mistral launches the Responses API
Built for building agents, the public preview returns the model's answer, reasoning, and tool calls as separate items, making multi-step workflows easier to handle than with Chat Completions.
Partnership
OpenAI and Oracle bring enterprise apps into ChatGPT
The deal brings Oracle applications and data into ChatGPT and adds new OpenAI-powered offerings for U.S. customers.
Developer tooling
Vercel AI Gateway introduces virtual models
A single virtual model ID centralizes model strings and configuration across agents, enabling instant model switching and enforced zero data retention.
Parsing
LlamaIndex releases Document Parsing Protocol
The DPP standardizes outputs from different OCR models and parsers on structures such as tables and dual headers, reducing agent disagreements in document understanding.
Speech
NVIDIA open-sources Nemotron 3.5 Transcribe
A 30B MoE with about 4B active parameters, 40-language support, up to one-hour recordings, promptable transcription and speaker labels, with day-one vLLM and SGLang support.
Tokenizers
H-Net learns to split code variable names
Tri Dao reports that H-Net dynamically splits variable names in code into separate tokens and scales better than BPE on code data, showing tokenizer design still has room to improve.
Paper
Hy-Nomas: an interactive video world model
An 8B diffusion-backbone world model uses causal attention and structured conditioning for multi-turn text instructions and continuous shot control in interactive video generation.
Robotics
ThinkWorld uses world imagination for VLA reasoning
Robot actions become discrete motion tokens in a Chain-of-Dynamics, with predicted futures aligned to real observations to use action-label-free human videos and improve cross-embodiment generalization.
Numerics
New paper flags flaws in standard softmax
Softmax Collapse rounds the target logit gradient to zero, while Softmax Cancellation produces nonzero gradients even when student and teacher logits are equal. The authors offer a concise fix.
Hardware
Build-A-Chip-Bench tests full-flow chip design
The benchmark takes AI agents from workload specs to GDS layouts, assessing end-to-end chip design with open-source tools and PDKs. NVIDIA contributed.
Coding agents
Opus 5.5 marks another turning point
Simon Willison calls October 2026 an inflection point for vibe coding: Claude Opus 5.5 has clearly improved in 3D, music, and game design, and can now build genuinely fun browser games.
Business
Anthropic S-1 reveals Opus 5.5 costs
The filing shows about $16.286B in first-half revenue and $2.380B in cost of revenue, implying roughly 85.4% gross margin and Opus 5.5 pricing around $0.6/$3.0.
Benchmarking
AA v5 index update sparks open-model debate
Natolambert argues any single score involves tradeoffs: AAv5 leans toward frontier general-agent ability, with eight closed models ahead of the best open-weight model, but it may undervalue open models on cost and control.
Vision
ImagiRep rethinks visual spatial reasoning
Across 922 questions and seven task types, accurately imagining RGB does not guarantee correct spatial reasoning; depth, panorama, and top-down representations add complementary gains.
Legal
Claude for Legal adds a tool library
Anthropic added ready-to-use Artifacts tools covering citation checks, contract review, and tracking new regulations.
Work
Kai-Fu Lee: AI will rewrite corporate hierarchy
In his FT op-ed, he predicts white-collar workers and middle managers who can collaborate effectively with neither AI nor people face the greatest replacement risk.
Infrastructure
Virtual models are "DNS for models"
Vercel's Guillermo Rauch describes the alias as resolving to a model with traffic-steering abilities, avoiding teams stuck on old versions and removing decision paralysis.
CI
Vercel CI targets the AI coding bottleneck
Fast, stable infrastructure plus Turborepo integration and agentic CI, operated across 20 worldwide clusters running millions of microVMs.
Partnership
Vercel to serve Particle models via AI Gateway
Particle optimizes open-weight models for token efficiency and serving cost; Vercel will surface its models and inference through AI Gateway.
Earnings
Alphabet hits a half-trillion run rate
Sundar Pichai: annualized revenue reached $500B, up 25% year over year, with AI investment pushing search query growth to a five-year high.
Speech
Voxtral ranks second on Voice Arena
Mistral says Voxtral Mini Transcribe 3 ranks second on the Diarization + ASR leaderboard.
Inference
SGLang serves Nemotron from day zero
Nemotron 3.5 Transcribe runs on SGLang-Omni through the /v1/audio/transcriptions endpoint with streaming and custom prompts.
AI results used in quantum Yang-Mills research
OpenAI shared a KITP working-group statement that a scientific collaboration is interpreting AI results on the existence of quantum Yang-Mills.
Runway Dev plugin comes to Codex
One plugin install lets Codex build Runway's API integration autonomously from a description of what the app should generate.
Replit proposes Agent Habitat architecture
Security boundaries, real failure experience, and a testable maturity scale for evaluating self-driving agent environments.
Kling 4.0 lands on Pika
Full reference, multiple keyframes, and up to 30-second generations; officials call it Kling's most controllable version yet.
Spark releases 1.7B on-device agent model
Spark-X2.6-1.7B supports 1M native context and controllable thinking, approaching the 4B version on SWE-Verified and τ²-bench.
Perplexity partners with Particle
Particle focuses on token efficiency and serving cost for open-weight models, working with Perplexity on search and model tools.
Perplexity rolls out new voice mode
Pro and Max users can handle parallel tasks by voice on web and Mac; officials call voice a more natural way to interact with computers.
Samaya releases finance research model Sterling
Post-trained for difficult financial research, it claims to beat frontier models on related tasks. Yann LeCun shared it.
vLLM teams ship Vela 3.0 router
Routing decisions and span extraction in one call, across six open models from 0.3B to 27B.
Ollama × FriendliAI Seoul meetup
A Nov 5 event on building production-grade agents with open models such as GLM, Kimi, DeepSeek, and MiniMax.
fx coding agent gets faster
The tiny, open, native CLI agent written in Zig is now faster, smaller, and more efficient.
Vidu Q4 Preview for ad teams
More creative variants with consistent product appearance, better lighting, and multilingual text rendering.