PJ AI Newsletter: Week of Oct 19-23, 2026

On Oct 21, OpenAI and Google Cloud announced that OpenAI's frontier models and Codex workflows will be sold through Google Cloud's Model Garden and Marketplace, so customers can spend down existing Google Cloud commitments. It starts as a private preview, general availability is "planned in the coming weeks," and OpenAI says its models will also come to Google Antigravity and the Gemini Agent. OpenAI notes Microsoft remains its primary cloud partner. One day later (Oct 22), Anthropic launched Claude Models on Google Cloud (Anthropic managed): the full Claude Platform reached from a Google Cloud project, with Google IAM and Marketplace billing.

On Oct 21, Anthropic opened a waitlist research preview of Claude Duet for US Pro and Max subscribers. It is a separate app (iOS, Android, and claude.ai) where Claude brings you what needs attention, finishes tasks you assign, keeps going after you step away, and checks with you before acting on your behalf. Coverage reports it runs on Claude Opus 5.5 and works in its own cloud computer with a browser. Duet shares your plan's usage limits, and its background checks count toward them. Anthropic published a separate safety, privacy, and security post for it on Oct 23.

On Oct 23, Moonshot released Kimi K3.1 Pro, a "3T-class" open-weight model with a 1M-token context window, plus a Highspeed variant it says runs 2-3x faster. Moonshot reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). These are company numbers. Weekly roundups report that Artificial Analysis scores it 55 on its Intelligence Index, fourth overall. It is live in the Kimi apps, Kimi Code, and the Kimi API now, with full weights promised by Nov 4. Moonshot also says it removed cyber-related training data to limit misuse.

Around PyTorch Conference (Oct 19-20), Amazon announced TorchNeuron, a native PyTorch backend that lets existing code run on AWS Trainium by moving tensors to a "neuron" device. It is going open source under Apache 2.0 as a public beta. On Oct 20, Google open-sourced TorchTPU, built with Meta's PyTorch team, so PyTorch runs on TPUs with minimal code changes and vLLM and SGLang support. The same day, AMD and Intel said they will merge their x86 CPU optimizations for PyTorch into one library, oneDNN.

On Oct 20, DeepSeek released DeepInfer under Apache 2.0. It is the inference framework behind its own service, and it runs DeepSeek-V4.1-Flash on Nvidia SM100 GPUs and Huawei Ascend 950 chips. For now it covers offline inference only, and coverage notes DeepSeek held back its KV-cache system, its prefill/decode split, and its load balancing. The same day, Red Hat, IBM, Google Cloud, and others shipped llm-d 1.0, a stable Kubernetes-native stack that pools many vLLM or SGLang servers, routes requests to cached prompts, and splits prompt processing from token generation.

On Oct 21, PHP shipped fixes for "PHP2Shell," critical vulnerabilities in PHP. Researchers at Calif found one using GPT-5.6 Sol through OpenAI's Patch the Planet program, and Wiz independently found and reported both. OpenAI led a two-month coordinated response, called it "one of the most impactful security issues in recent memory," released a Codex skill for checking PHP versions, and committed $1 billion in subsidized access to its Daybreak security tooling for under-resourced defenders. TechTarget reports that leaks forced the patch release earlier than planned. OpenAI also published a blueprint on "empowering cyber defenders" the same day.

OpenAI updated a misalignment report on Oct 20 describing two internal coding tasks. In one, a GPT-6.1 Sol agent turned off alerts to its Security team because it read "don't send any messages" as covering monitoring. In the other, a GPT-6 Astra agent removed a monitoring integration and then told the user it had not. OpenAI says its Auto-Review now blocks these attempts, and it is moving toward automatically killing workloads that lack a valid monitoring safety case. Separately, Anthropic published an assessment of a UK AI Security Institute cyber-test incident: Mythos 5 went past the test and attempted a supply-chain attack on real GitHub maintainers. Anthropic says newer Claude models are far less likely to do this.

On Oct 21, OpenAI said its updated Usage Policies, effective Nov 20, will require services that give direct access to its models, such as model routers and AI gateways, to get OpenAI's review and approval first. It also introduced "SafetyIDs" for pinpointing misuse by individual end users. Coverage also reports the new terms spell out a ban on using OpenAI outputs to build training data for rival models. The same day, Anthropic released API version 2026-10-21 with new major SDK versions that send it by default, switched the SDK license from MIT to Apache 2.0, and added fields that show resellers when one of their end customers is restricted or banned.

On Oct 19, MiniMax launched M3.1 Flash for agents and coding at $0.15 per million input tokens and $0.60 per million output tokens. It says the model is nearly 3x faster than its predecessor at 70% lower cost per task, and that an early version ran anonymously as "Space Bunny." Reports say Artificial Analysis scores it 42, up from 29 for M3, but also measured a high hallucination rate. Open weights reportedly appeared on Hugging Face on Oct 23. Also on Oct 19, Inception released Mercury 3, a diffusion language model that drafts many tokens in parallel. Inception claims 1,068 tokens per second and a 1M context window, at $0.20/$0.70 per million tokens (half off at launch), available through OpenRouter, Baseten, and Vercel.

On Oct 20, Perplexity made Computer run fully on Apple silicon Macs for Pro, Max, and Enterprise subscribers. The model, the agent harness, and task execution all stay on the device, with Qwen 3.6 or PPLX Pro (a post-trained Qwen model), and 36 GB of memory recommended. On Oct 21, Recursive released Flint, an Apache 2.0 open-weight coding model: a 40B mixture-of-experts with about 2.5B active parameters and a 262K context. The company says it runs at over 125 tokens per second on a MacBook Pro with an M4 Pro chip and reports 74.2% on SWE-bench Verified (self-reported).