Top Story · Cloud & Agents
OpenAI frontier models land on Google Cloud
OpenAI brings frontier models and Codex agent workflows to Google Cloud Model Garden, letting enterprises invoke them in existing cloud environments.
OpenAI and Google Cloud are converging on the enterprise. The partnership places OpenAI's frontier models — and its Codex agentic workflows — inside Model Garden, Google Cloud's curated model catalog, so companies can call the technology from the cloud where they already build. It is a pragmatic answer to a market that increasingly buys AI where its data and infrastructure already live, rather than routing work to a separate platform. For Google, the deal puts a top-tier model family alongside its own Gemini; for OpenAI, it opens a distribution channel measured in enterprise data centers rather than individual subscriptions.
NVIDIA DSX POD cuts AI factory delivery to 14 days
At GTC Berlin, NVIDIA introduced DSX POD, a pre-integrated, pre-tested AI factory that can be deployed at a prepared site in as little as 14 days — a sharp contraction from the months data centers typically require. The pitch is aimed at AI builders who need computing capacity they can put to work sooner rather than later.
Security
OpenAI uses AI to find major PHP vulnerability
OpenAI's Patch the Planet initiative used its models to help researchers discover critical security flaws in PHP, the open-source language behind an estimated 70% of websites. Over two months, the program led a coalition of maintainers and researchers through disclosure and patching.
The follow-through shipped as a packaged Codex skill, php2shell-checker, that lets developers check whether a website is affected using only HTTP headers or the version exposed by php -v.
Model Release
Flint 40B MoE local coding model released
Recursive launched Flint, a 40B mixture-of-experts model with 2.5B active parameters and a 262K-token context window, released under Apache 2.0 and aimed at local agentic coding. It runs in roughly 24GB, making it feasible on a single workstation rather than a data center.
Research
Perplexity proposes JOLT self-distillation
JOLT is a post-training method in which one model acts as both teacher and student — the teacher is trained to give feedback the student can actually use. On the MATH benchmark it matched GRPO accuracy while using 6.5x fewer training tokens.
Inference
vLLM native PyTorch inference comes to TPU
Built with Google and Inferact, vLLM-TorchTPU brings native PyTorch serving to TPUs, adding prefix caching, speculative decoding and disaggregated serving to the accelerator so existing PyTorch tooling translates directly.
Industry · Partnership
Mistral and NVIDIA advance industrial AI
Mistral and NVIDIA are co-developing AI-native engineering workflows focused on three fronts: faster simulation, open benchmarks and datasets for physics AI, and evaluations for agentic AI in CAE and EDA. The collaboration targets the manufacturing world, where simulation cycles and design iteration are the bottleneck.
"AI for industrial engineering will bring humanity to its next technological era."
Arthur Mensch · Mistral AI
Drug Discovery
GPT-Rosalind finds new neurodegenerative clue
GPT-Rosalind surfaced a new lead for drug discovery: could restoring cholesterol transport protect neurons from degeneration? The finding links DCTN1 — a gene tied to Perry syndrome, a rare neurodegenerative disorder — to changes in mouse-brain data, suggesting a fresh angle on neuron protection.
Open Source
vLLM open-sources Humming quantization kernels
vLLM and Ant Group open-sourced Humming, a single kernel library unifying multiple quantization formats and GPU workloads. The project reports kernel speedups of up to 5x and a 14–43% improvement in end-to-end MoE throughput across the workloads it covers.
Platform
Claude API releases its first formal version
The Claude API launched version 2026-10-21, its first formal API version release since 2023. Anthropic plans more frequent versions so developers can handle breaking changes before a new model forces them, while the older 2023-06-01 version remains supported.
Product
ChatGPT ad stack introduces LLM orchestration
OpenAI explained how ChatGPT's advertising stack uses an LLM orchestrator, retrieval, ranking and auction tools working together — while keeping answers independent of the ads served alongside them.
Agent Security
NVIDIA releases OpenShell agent security architecture
The OpenShell reference architecture sets boundaries for enterprise AI agents and stops them when they step outside those boundaries, giving businesses a reference design for securing agents at scale.
Search
Cohere releases Rerank 5
Cohere launched Rerank 5, advancing AI retrieval and ranking precision, available in Pro and Fast tiers — a few lines of code that change how search results are ordered.
Spark-X2.6-4B supports million-token context
Released under Apache 2.0 with 1M-token context and controllable thinking, supported by vLLM on day one.
Gemini 4 Argon generates interactive anatomy
Three open datasets combined into a 3D human anatomy view with toggleable layers, isolated muscles and a zoomable heart.
MiniMax M3.1 Flash joins voice agents
Combined with SGLang Omni for real-time voice agents that can be interrupted, corrected and redirected.
NVIDIA opens simulation assets to Hugging Face
SimReady Central Library brings 2,000+ simulation-ready 3D assets to Hugging Face for robotics and physical AI.
NVIDIA and partners expand AI compute
Working with Amkor, Samsung Electro-Mechanics, SPIL and Unimicron on components, advanced packaging and testing.
NVIDIA Isaac tools support robot agents
New Isaac tools help developers train, test and deploy robot agents; Agile Robots also showed new products.
Kling 4.0 API officially opens
The video generation API opens to products and workflows, letting developers integrate Kling 4.0 directly.
Realtime open video generation accelerates
FastH3-Omni ships with weights and training code; vLLM-Omni serves it day zero, including on consumer RTX GPUs.
OpenAI releases PHP2Shell check skill
A Codex skill that checks whether a site is affected using HTTP headers or the php -v version.
Annual public benefit report released
Five pillars and the actions advancing responsible advanced AI, published under its PBC charter.
Anthropic Trust welcomes a physicist
Harvard experimental physicist Christopher Stubbs joins the Long-Term Benefit Trust.
Papers with Code PDF assistant enhanced
Upgraded to a full RAG pipeline with citable sources, improving paper reading and retrieval.
Sentence Transformers releases 6.2.0
Adds mergekit model merging, faster loading and an updated embedding toolchain.
Runway offers unlimited Kling 4
Three days of unlimited access for new Max and Team plans, plus 25% off across plans.
v0 integrates Slack MCP
Read threads and files, post updates back to Slack, limited to channels the account can see.
Replit integrates Stripe MCP to register companies
Replit Agent can help users incorporate a business through the Stripe Atlas MCP integration.
Ideogram 4.5 strengthens real edits
Focused on real editing scenarios: fixing anatomy, spelling errors and artifact buildup across multi-turn edits.
Adobe Firefly integrates Kling 4.0
Kling 4.0 is now available inside Firefly, letting users invoke video generation directly in the studio.
vLLM matches DeepSeek decode within 6%
With Wide-EP and EPLB, vLLM matches DeepInfer's decode performance within 6% at roughly $39 per GPU-hour on the AgentX workload.
SGLang Omni unifies full-duplex voice agents
A unified runtime for full-duplex interactive agents, with streaming and playback-aware interruption handling for barge-in.
SGLang has day-zero support for Flint
Flint, a 40B MoE built for fast local agentic coding, is served by SGLang from day zero.
SGLang-TorchTPU powers Meta inference
A TPU plugin built with Google and Meta's PyTorch team, now powering production inference at Meta.
Vercel ships Firewall mitigation for Node.js RCE
Proactive Firewall-level mitigations rolled out ahead of the official patch, protecting all deployments.
DeepSeek v4.1 Flash joins Vercel Token Pass
Fast, cheap and served with zero-data retention, pitched as a strong subagent for executing larger models' plans.
Industry moved weeks early to blunt PHP2Shell
Weeks of quiet coordination ahead of disclosure, with WAF mitigations already blocking attempts on PHP origins.
durable-harness keeps agent sessions alive
Gives Pi, OpenCode, Claude Code and Codex persistence on Vercel, Postgres or a local machine across crashes and timeouts.
Perplexity adds /graphify to Computer
A skill that visualizes a codebase, with 5,000 credits given to every user to try it.
JOLT research detailed with NVIDIA AI
Self-distillation plus RL post-training, in which the teacher gets extra information the student does not.
PHP patches two AI-found flaws
Emergency releases 8.5.12, 8.4.27, 8.3.36 and 8.2.35 close two unauthenticated server-takeover bugs.
Meta runs AI inference on Google TPU
Google published a TPU version of SGLang, citing Meta's production use for inference.
Codex cache becomes permanent
Logged-in users resume old sessions with guaranteed cache hits and no miss penalty.
Claude Duet reaches out proactively
A new assistant that initiates contact; US Pro and Max users join a waitlist at no extra charge.
"How Codex Happened" dissected
Every's detailed account traces how the Codex App was driven into existence.
Dario answers on safety vs superintelligence
Accepts three communication missteps but keeps course after Encode's questioning.
Jev: the AI that only fills bubbles
A model that does one thing — choosing among prepared options, like marking an exam sheet.
Four companies, one product category
Grok Bot, Muse, Dots and a new Anthropic assistant all landed within just over two months.
AI semiconductor endgame, part four
Whether cheap open-source models dent OpenAI and Anthropic's GPU-buying economics.