AI DAILY / DEV
MONDAY
November 16, 2026

    Google Ships Gemini 4 Hydrogen — $1.50 Input, 2.5x Faster Than 3.8 Flash

    • Nov 12 launch: workhorse tier beneath Gemini 4 Carbon and Argon, pitched at everyday agentic coding.
    • Pricing $1.50 / $7.50 per million input/output tokens with a 1M-token context window.
    • 57% on Terminal-Bench 4.1 vs 18.2% for Gemini 3.8 Flash; 51 on Artificial Analysis Intelligence Index at default thinking, 54 at xhigh.
    • Google claims 81% fewer agentic overreach cases and 5x fewer fabricated claims vs 3.8 Flash.
    • Rolling out across Gemini API, AI Studio, Gemini for business, the Gemini app, and Search.
    models marktechpost.com

    WSJ: Nvidia Allegedly Engineered GPU Scarcity by Funding Its Own Buyers

    • TechPowerUp write-up of the WSJ report tops Hacker News at 1,051 points, 624 comments.
    • Nvidia reportedly in talks to provide billions in guarantees to startups that then buy its chips and build data centers.
    • Company filings confirm customer financing and credit support, but critics argue the scarcity framing isn't proven.
    • Lands days before Nvidia's first post-Jensen earnings and Anthropic's IPO week — compounds the circular-financing debate.
    industry news.ycombinator.com

    llama.cpp and ggml Hit v1.0.0 — Semver Arrives for Local AI After 3.5 Years

    • ggml-org shipped both libraries on Nov 13 with a hard switch to semantic versioning for the local inference stack.
    • 35 deprecated functions removed from the public C API; NextN API folded into llama.h, llama_set_warmup gone.
    • ggml 1.0.0 drops legacy DIAG_MASK ops, adds configurable YaRN truncation and FP8 KV-cache support.
    • Distributed via 8 channels — Debian, Homebrew, conda-forge, Winget, Docker, and more.
    • Discussed on HN at 243 points, 78 comments.
    open-source github.com

    Thinking Machines Opens Inkling 2 Small Weights — 594B MoE, Apache 2.0

    • Nov 12 drop on Hugging Face under Apache 2.0, one week after the hosted-only Tinker debut.
    • 594B parameters with 17B active at inference, 1M-token context, image input.
    • Scores 31 on the Artificial Analysis Intelligence Index; model card reports 30.6.
    • Safety evals cover sycophancy, CBRN, cyber, and loss-of-control risks; lab concludes no material uplift beyond the open-weight ecosystem.
    • Positioned as the strongest US open-weight model for FrontierCode; still trails on Terminal-Bench.
    open-source alphasignal.ai

    Cloudflare Open-Sources Kitesurf, Its Agent Browser Engine

    • Changelog dated Nov 12, blog post Nov 13; Apache 2.0 on GitHub.
    • Rust engine compiled to WebAssembly — Blitz for HTML/DOM/layout, Stylo for CSS, Boa as JS fallback.
    • Parses HTML, runs page JavaScript, and renders screenshots on Workers with no browser container — Cloudflare cites 3–7x lower memory/CPU vs Chromium for agent workloads.
    • Self-hosting still requires a Workers Paid plan to use Dynamic Workers for per-page isolation.
    • Cloudflare flags it as experimental with incomplete Chrome compatibility.
    open-source cloudflare.com

    Anthropic Opens Private Beta for Self-Hosted Claude Managed Agents

    • Nov 12 beta: the entire orchestration stack — harness, sandboxes, sessions, credentials — runs in the customer's cloud or on-prem.
    • Customers operate the Managed Agents server, its database, and workers that execute shell and file tools.
    • Model calls can route to Anthropic, Amazon Bedrock, Google Agent Platform, Microsoft Foundry, or an internal gateway.
    • Credentials stay in the customer's secrets manager; a proxy outside the sandbox attaches them to MCP and API calls.
    • Access gated through Anthropic account teams — no self-serve enrollment, no GA date announced.
    tools helpnetsecurity.com

    OpenAI Disbands Standalone Robotics Team as General Models Take Over

    • Fortune exclusive, Nov 13: researchers fold into the main research org under CRO Mark Chen; applied and hardware staff move into Applied under Greg Brockman.
    • Second time OpenAI has broken up a robotics group — the first wound down in 2020 citing limited data.
    • Standalone team was formed in 2024 alongside the Figure AI partnership.
    • OpenAI still developing an in-house humanoid; Altman targets a first public demo in 2027.
    industry fortune.com