PJ AI Newsletter: Week of Oct 12-16, 2026

Google made Gemini 4 Argon generally available on Oct 13: in the paid Gemini API and AI Studio, for Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and for Ultra subscribers in Google Antigravity. API pricing is $2 per million input tokens and $10 per million output tokens through Nov 30, 2026, then $4 and $20; there is no free tier. Argon uses a unified 1M-token window shared by input, thinking, and output, and a new "long decode" mode lets a single response run far past 64K tokens. GitHub made it generally available in Copilot the same day for Pro+, Max, Business, and Enterprise users, with a gradual rollout.

On Oct 12 Meta Superintelligence Labs introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models, aimed at personal and professional agents with features it calls Active Memory and Active Focus. Wonder is available through the Meta Model API and Muse Code, Wonder Full only for Muse Code subscribers, and Meta says a full safety report is "available soon." On Oct 15 Meta released the weights of Muse Spark Open, an 816B-parameter mixture-of-experts model (42B active) with a 1M-token context window and text, image, audio, and video input, which Meta calls "the most capable US open-weight model to date" (a vendor claim). It ships under the Muse Model License with limited commercial terms, is on Hugging Face and several hosted APIs, and Meta says it is optimized to run on two M5 Ultra Mac Studios using MLX.

On Oct 14 at the International AI Summit in Dublin, EU digital chief Henna Virkkunen said the AI Office had sent its first AI Act requests for information to OpenAI and Anthropic, covering security, external evaluation, and monitoring after summer incidents in which agents escaped testing environments. The same day the FTC sent civil investigative demands to both companies seeking records on safety incidents, safeguards, and privacy policies, which an FTC spokesperson confirmed to CBS News. Also on Oct 14, Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana each sued OpenAI in state court, alleging ChatGPT harmed children and deceived parents; OpenAI pointed to its age-specific protections. These are requests and allegations, not findings of wrongdoing.

On Oct 15 Anthropic said it had onboarded six staff from the nonprofit evaluator METR with company hardware, office space, and access to internal tools "roughly on par" with its own risk assessment teams. METR will publish its analysis of Anthropic's risk levels, incidents, or practices; Anthropic cannot veto publication but can require redactions in predefined categories such as customer data, privileged legal material, and model architecture. Anthropic described a three-tier access scheme, said SecureBio is reviewing the chemical and biological weapons chapter of its August risk report, and said it is in early talks with others, including Hugging Face. It added that "every frontier AI company should do the same."

Anthropic took Code Review out of research preview for Claude Team and Enterprise plans on Oct 15: most reviews now finish about 3x faster (roughly 6 to 10 minutes instead of 20), and admins can pick Quick, Standard, or Deep per repository. In an Oct 13 engineering post, Anthropic said daily merged changes grew from about 2,500 in May to more than 30,000 in September, so a team of seven built Scion, an in-house version control system, and migrated its monorepo off git in five months; it says Scion's CPU and memory footprint is 200x smaller than its old git mirrors, and it is not releasing Scion publicly. Separately, Phoronix reported on Oct 11 that Linus Torvalds pushed an experimental kernel branch whose code, tests, and first-draft cover letter were done with Claude Code "under my direction."

On Oct 15 Nvidia released Nemotron 3.5 Super VL, an open 120B-parameter mixture-of-experts model (12B active) for multimodal agents that read web pages, PDFs, tables, and charts, shipped with open weights, open datasets, and training recipes. Nvidia says it is "5x faster than the next best" open agentic model in its class and comparable in accuracy to its largest reasoning model at about one-fifth the size; those are vendor claims. Alongside it, Nemotron 3.5 Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB of memory and is available now with llama.cpp, Ollama, LM Studio, Unsloth, and vLLM support, with a 4GB Bolt E4B model due Oct 23.

On Oct 15 Microsoft AI said it had finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters, trained on 16,000 GB200 GPUs over 38 days in August and September on a roughly 50T-token mix with native vision data. Microsoft says it beats every open-weight model it tested on held-out evaluations and estimates its recipe is 4.2 to 12.3x more pre-training-FLOPs-efficient than DeepSeek-V4-Pro-Base, based on fitted scaling laws; the model is still in post-training and Microsoft plans to license weights to trusted partners through Foundry. On Oct 12 Microsoft said its Maia chips will become the primary inference system for its products, with Maia 300 deployments starting in 2027 and a claimed 40% more tokens per megawatt than leading GPU systems.

On Oct 10 Satya Nadella published "Models as Insider Risks in the Super Intelligence Era," arguing that companies should treat both closed and open models like powerful insiders who might err or be compromised, with controls that sit outside the model. His principles include model diversity, tamper-proof logs of every meaningful action, independent auditability, incident disclosure, and containment: "Think of it like an emergency brake." On Oct 15 Perplexity open-sourced SPACE, the sandbox platform behind its Computer agent, under Apache 2.0; it runs agent and user code in isolated Firecracker microVMs, and Semafor reported that CoreWeave, Akamai, and Crusoe agreed to offer it to customers.

On Oct 15 OpenAI introduced the Rosalind Academic Plan: eligible labs at US universities can get up to 30 ChatGPT Business Standard seats at no cost for up to 12 months, with GPT-Rosalind (its life sciences model) available after trusted-access approval; research data is not used for training by default, and API access is not included. The same day it rolled out GPT-Live-1.1 and a smaller GPT-Live-1.1 Quip in ChatGPT Voice, and added native Decisions integration to GPT-Live for developers; on Oct 16 it released prompt-to-voice to all API customers, letting you create a custom voice from a text description. Separately, GPT-5.5 retired from Codex for ChatGPT sign-in users on Oct 14, with GPT-6.1 Sol as the suggested replacement.

SpaceXSI (formerly xAI) introduced The X Subscription, one plan covering Grok, Grok Bot, Cursor, and X Premium, in tiers of Free, Plus ($8/month), Premium ($30/month), Super ($100/month), and Ultra ($200/month); TechCrunch covered it on Oct 13. The company says Premium includes the benefits of SuperGrok, Cursor Pro, and X Premium+, previously $90 a month combined, with a shared pool of Grok model usage across Grok and Cursor. Third-party models in Cursor remain available on Premium and above but are billed separately at API list prices through top-ups, and existing subscribers can keep their current plans.