PJ AI Newsletter: Week of Nov 9-13, 2026

On Nov 9, Anthropic upgraded its top tier with Claude Fable 5.2 (available to everyone) and Claude Mythos 5.2 (the same underlying model, offered only to organizations verified through its cyber and life-sciences programs). Anthropic says the models output up to 30% faster than the 5.1 versions and cost $8 per million input tokens and $40 per million output (down from $10 and $50), with a 1M-token context, 128k max output, and an optional fast mode up to 2.5x faster at double the price. On Anthropic's own table, Fable 5.2 scores 49.1% on FrontierCode 2.0 (xhigh effort) versus 45.5% for Opus 5.5 and 69.9% on FrontierSWE v2. Anthropic still recommends Opus 5.5 as the daily driver, with Fable as an Advisor or Orchestrator for the hardest work. API notes flag breaking changes: a stricter prefix check on thinking blocks and a new computer-use toolset.

On Nov 10, Google made Gemini 4 Carbon, its strongest model for long-horizon coding, science, and cyber work, available to paying API developers, enterprises, and Google AI Ultra subscribers, at an introductory $2/$10 per million tokens through Nov 30 (then $4/$20). Two days later it launched Gemini 4 Hydrogen, a fast model for everyday coding and agent work, rolling out globally in the Gemini API (all tiers), AI Studio, the Gemini app, Google Search, and Antigravity for paid subscribers. Google says Hydrogen beats Gemini 4 Argon on the Artificial Analysis Intelligence Index (51 vs. 50 at default thinking, 54 at xhigh) and costs a flat $1.50/$7.50 per million tokens across its 1M context. 9to5Google reports Google's claims of more than tripling Terminal-Bench 4.1 versus Gemini 3.8 Flash (57% vs. 18.2%) and finishing tasks 2.5x faster. Google also called Hydrogen the biggest model upgrade to Search yet.

At its Cyber at the Frontier event on Nov 12, OpenAI introduced two Daybreak products. Daybreak Red Team uses multiple agents to pentest web applications, using source code and app context to investigate attack paths and confirm vulnerabilities, inside isolated environments with Guardian review enforcing testing boundaries. Defense Factory traces findings (from Red Team or tools like Wiz, Snyk, and Endor Labs) through your code and prepares patches, which engineers review, edit with Codex, and open as draft GitHub pull requests; nothing deploys automatically. Red Team is in early access, and Defense Factory is a preview plugin in ChatGPT for Daybreak workspaces. This follows last week's GPT-6 Cyber launch.

On Nov 12, Irreplaceable Education Fund, Public Citizen, Heat Initiative, and The Alliance for Secure AI sued Anthropic, Google, Meta, OpenAI, and SpaceXSI in San Francisco Superior Court. Invoking California's Unfair Competition Law, they ask the court to freeze development of models more capable than today's, citing the companies' own warnings about cyberattacks, bioweapons, and loss of human control. The complaint was filed by former New Jersey Attorney General Matthew Platkin. It is a filing, not a ruling: no court has ordered anything.

On Nov 12, Nvidia's board announced that Jensen Huang resigned as president, CEO, and board member, effective immediately, after 33 years leading the company. CFO Colette Kress is interim CEO while the board searches for a permanent successor, and Harvey C. Jones (former Synopsys CEO) becomes the board's first-ever chair. Nvidia says operations, customer commitments, and the product roadmap continue without interruption. The announcement gave no reason for the resignation, so treat any explanation you read elsewhere as unconfirmed.

On Nov 12, Thinking Machines published the weights for Inkling 2 Small on Hugging Face under Apache 2.0, after a week of closed access on its Tinker platform for safety testing. It is a mixture-of-experts model with 594B total and 17B active parameters, supports image input and variable effort, and scores 31 on the Artificial Analysis Intelligence Index per the company. It is integrated with vLLM, SGLang, and Unsloth, and offered via API at providers including NVIDIA, Fireworks, Together, and Cloudflare. The catch: the main NVFP4 checkpoint needs roughly 500 GB of aggregate GPU memory.

On Nov 9, Meta Superintelligence Labs released Muse Ember, its smallest and most cost-efficient model, built for tool calls, coding subtasks, routing, extraction, and long document parsing. It has a 1M-token context, handles text, images, and audio, and starts at $0.10 per million input tokens and $0.50 per million output (rates triple above 200K input tokens). Meta pitches it as the volume worker under its frontier Muse Wonder: "Muse Wonder plans, Muse Ember handles the volume." Meta's cost-per-task comparisons against Haiku and Luna are its own claims.

On Nov 12, Anthropic launched self-hosted Claude Managed Agents in private beta. Companies can run the entire Managed Agents service on their own infrastructure, not just the sandboxes, and keep agents, sessions, and session history there. Model requests go to the provider the customer chooses (Anthropic by default). Access is by request through Anthropic's account team.

Anthropic moved Claude deeper into team workspaces this week. On Nov 10 it launched Claude Whiteboards and Claude Sheets in beta and made Claude Dashboards generally available as Claude Data (with certified queries and live sharing), on all paid plans, and let interactive connectors open as full apps beside the chat. On Nov 12, Claude Meetings entered beta on Team and Enterprise plans: Claude joins Google Meet calls as a visible participant, answers questions, builds docs and whiteboards live, and writes notes and action items. By default it stays out of calls with outside guests, and it saves transcripts and summaries, not audio.

On Nov 11, Google introduced the Gemini Decisions API, powered by a new Gemini Decision 1 model. Instead of generating text, it scores candidate options for closed-set questions (yes/no, multiple choice, rubric scores) in a single forward pass over one multimodal input of up to 256K tokens, returning probabilities and confidence. Google cites 150 to 250 ms end-to-end latency, $0.10 per million input tokens, and no output-token charges, with a no-cost tier in AI Studio. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, and others.