NPDev Daily AI Briefing — November 13, 2026
Nvidia lost its CEO of 33 years: Jensen Huang resigned and CFO Colette Kress is interim CEO, with Q3 earnings still set for November 17. The same 24 hours brought two cheap-and-fast model releases — Google's Gemini 4 Hydrogen and OpenAI's GPT-6.1 Terra — plus llama.cpp 1.0.0, the first API/ABI stability promise for local inference.
Written by NPDev Labs with AI assistance and reviewed before publication. How we work.
Nvidia's board accepts Jensen Huang's resignation, names CFO interim CEO
Nvidia's board accepted Jensen Huang's resignation as president, CEO and director effective immediately, the company said on November 12; the 8-K filing records the board action on November 11. CFO Colette Kress, at Nvidia since 2013, is interim CEO while keeping the CFO role, and Harvey C. Jones becomes the board's first-ever chair. The company says Q3 fiscal 2027 results will still be reported on November 17 as scheduled.
Why it matters: Nvidia's product roadmap and allocation decisions drive GPU supply for every serving stack. The announcement names no reason and gives no guidance change, so treat the roadmap as intact until the November 17 call, and treat capacity assumptions as a risk to hedge rather than a settled input.
Context and caveats: Reported, not confirmed by Nvidia: AP and The Verge attribute the exit to disclosures about extramarital affairs, including one with a former Megaspeed executive suspected of being a Chinese government agent, and AP reported the stock fell 14% at Thursday's open. Nvidia's own statement gives no cause and thanks Huang for 33 years; the board chair said operations, customer commitments and the product roadmap should continue without interruption.
Sources: Read source: Nvidia newsroom — | 8-K filing (SEC) — | Reported context: AP —
Published weights point to export-controlled Blackwell chips in six Chinese frontier models
Researcher Will DePue (Unemployment Labs) published a weight-forensics analysis on November 12 finding that six open-weight models released since April — DeepSeek-V4-Flash and V4-Pro, Moonshot's Kimi K3 and K3.1-Pro, MiniMax M3/M3.1-Flash and Ant Group's Ling-3.1-flash — carry Philox initialization traces matching Blackwell part configurations. For DeepSeek-V4-Flash he reports 445 of 469 selected weight rows reproducing exactly on a B300 and none on an H100.
Why it matters: The method exploits thread counts: Hopper H100/H200 use 132 SMs, B200/B300 use 148, GB200/GB300 use 152, so a correct seed with the wrong thread count puts the right numbers in the wrong places. The finding establishes initialization hardware, not where training ran or how the chips were obtained — but it makes published weights an audit surface for compute access, which matters if you benchmark or procure Chinese open models.
Context and caveats: Confidence intervals of roughly 49 to 1,385 sigma are the researcher's own statistics and have not been independently replicated. The technique needs published weights, so it says nothing about closed Chinese models, and the analysis explicitly does not link its findings to the Nvidia leadership story.
Sources: Read source: analysis thread — | Reported summary: Moneycontrol —
Gemini 4 Hydrogen ships everywhere as Google's execution-tier model
Google released Gemini 4 Hydrogen on November 12 as the fast tier beneath Gemini 4 Carbon and Argon, available in the Gemini API, all Google AI Studio tiers, the Gemini app, Google Search, Gemini for business, and Antigravity for AI Pro and Ultra subscribers. Pricing is flat across a 1M-token window at $1.50 per million input and $7.50 per million output tokens. Google puts it at 51 on the Artificial Analysis Intelligence Index at default high thinking (54 at xhigh), above Argon's 50.
Why it matters: This is the tier most production agent loops will actually call: Google claims 2.5x faster completion than Gemini 3.8 Flash, Terminal-Bench 4.1 at 57% versus 18.2%, and more than 5x fewer fabricated claims. It also ships Orchestrator in Antigravity, which switches between Carbon and Hydrogen inside a session and, on Terminal-Bench 4.0, reached near-Carbon success with 24% less quota and 19% faster completion.
Context and caveats: Every number above is a vendor claim; the Intelligence Index figure is Google citing Artificial Analysis. Hydrogen's Terminal-Bench 4.1 score is not comparable to the Terminal-Bench 4.0 numbers OpenAI published the same day.
Sources: Read source: Google announcement — | Developer guide — | Search rollout —
GPT-6.1 Terra cuts price 50% and leans on computer use
OpenAI introduced GPT-6.1 Terra as the successor to GPT-5.6 Terra, priced at $1 per million input tokens, $0.05 cached input and $5 output — half the standard price of both GPT-5.6 Terra and GPT-6.1 Sol. It is available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and in the API as gpt-6.1-terra, with a 1.05M-token context window and 128K max output. OpenAI positions it for visual reasoning and precise computer use rather than hard coding.
Why it matters: The interesting claim is factuality: on OpenAI's internal error-inducing prompt set the error rate fell from 53.6% to 16.7% at low reasoning effort, and to 4.3% at maximum. Terminal-Bench 4.0 rose to 50.3% at max effort versus 21.5% for GPT-5.6 Terra. For high-volume document and UI automation, that combination of price and error rate is the practical argument; OpenAI still points at Sol and Astra for the hardest coding.
Context and caveats: OpenAI's safety addendum rates the model Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework, with the GPT-6 Astra safeguard stack. Benchmarks are OpenAI-run; the hallucination figure is described by independent evaluators as a large drop on their own benchmark, which is a different measurement.
Sources: Read source: OpenAI announcement — | Model docs and pricing — | Safety addendum —
Anthropic lets enterprises host the whole Claude Managed Agents service
Anthropic opened self-hosted Claude Managed Agents as a private beta on November 12, moving the service itself — API, agent loop, and storage for agents, sessions and history — onto customer infrastructure, not just the execution sandboxes. Customers run the server, its PostgreSQL 16+ database and their workers, and install each release themselves. The model still runs at Anthropic or at a chosen provider: Bedrock, Claude on Google Cloud, Microsoft Foundry or a customer gateway. Access is requested through an Anthropic account team.
Why it matters: It separates control-plane custody from inference: session history and credentials can stay in your network while model calls still leave with the accumulated conversation, including tool inputs and outputs. The docs also warn the shell tool is a real shell with the worker's permissions, and that some capabilities remain hosted-only — so read the differences page before assuming parity.
Context and caveats: Lineup context: Anthropic's newest models are Claude Fable 5.2 and Claude Mythos 5.2 (November 9); Claude Opus 5.5 (September 22) is still the model third parties benchmark against, including in OpenAI's Terra comparisons.
Sources: Read source: docs — | Anthropic overview —
llama.cpp hits 1.0.0 and commits to semantic versioning
llama.cpp and ggml shipped v1.0.0 on November 13, 3.5 years into the project. The release is mostly a contract: from here, a major-version bump signals a breaking change to the public C API in include/, minors add models and backends, and patches fix bugs without API changes. The project also published packaging channels — pre-release artifacts, Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework — and says both libraries will track the same versioning rules.
Why it matters: libllama and libggml are now dependencies you can pin like any other system library, which matters if you ship an on-device product or a runtime wrapper. The ABI promise is the real news: previously every weekly build could move interfaces, which is why so many wrappers pinned arbitrary commit hashes.
Context and caveats: No model or capability change is claimed in the post; the deliverable is interface stability and distribution.
Sources: Read source: release post — | GitHub release v1.0.0 —
Edge0 puts a 145B agent model on a single consumer GPU
Edge0 published 4-bit releases of Edge0-145B and Edge0-120B, latent sparse mixture-of-experts models with 145B total and 10.12B active parameters, 256K context and a 512-expert pool trained natively at 4 bits. Its model card reports 23.25 GiB VRAM plus 38.27 GiB host RAM on a 24 GB RTX 5090, 98.3 tokens per second agent decode, and 86.5 on Terminal-Bench 2.1, with an inference engine and plugin tooling to run it.
Why it matters: Tool-calling agents at 100B-class quality on one consumer GPU change the cost math for private, offline and batch agent work: no per-token bill, no data leaving the machine. The catch is memory residency — experts stream from host RAM — so throughput depends on how many experts fit in VRAM on your card.
Context and caveats: Read the licence: it is a custom Edge0 Community Model 1.0, not Apache 2.0 or MIT, despite some coverage calling the release open source, and the Hugging Face metadata lists Qwen3.8-Flash-Next as the base model. All throughput and benchmark figures are vendor-run on vendor-selected hardware and need reproduction.
Sources: Read source: model card — | Open-video demos and code —
IBM's Granite team publishes LLMON, a markup layer aimed at prompt injection
IBM's Granite team published the design write-up for LLMON (LLM Object Notation) on November 12, together with Granite 4.1 3B and 8B preview models trained to consume it. Instead of JSON, LLMON wraps semantically tagged components in reserved special-token delimiters, states schemas unambiguously for a causal model, and uses IDs to bind each instruction to the exact data payload it may act on. Converters turn existing JSON into LLMON, and IBM's Mellea framework applies it automatically.
Why it matters: It attacks prompt injection structurally rather than statistically: hidden instructions in retrieved documents are mostly ignored because instruction and data occupy explicitly different regions, and IBM reports attack success rates dropping to nearly zero on some benchmarks while holding general capability. That is a different lever from fine-tuning or classifiers, and it is cheap enough to test on a small model.
Context and caveats: Status check: the preview weights were already on Hugging Face before this post — the 8B repository dates to October 1 — so November 12 brought documentation and tooling, not a new release. The models are labelled experimental, and the injection results are IBM's own.
Sources: Read source: IBM write-up — | Model card: granite-4.1-8b-llmon-preview —
Practical implications for engineers
- Route by cost tier instead of defaulting to the frontier. Gemini 4 Hydrogen at $1.50/$7.50 and GPT-6.1 Terra at $1/$5 both claim near-frontier agent scores at a fraction of prior prices; benchmark them on your own eval set before switching, because both sets of numbers are vendor-run.
- Re-run hallucination and instruction-following evals if you use small or cheap models for extraction. OpenAI's error-rate drop from 53.6% to 16.7% on error-inducing prompts is the kind of change that invalidates a guardrail you added two model generations ago.
- Pin llama.cpp now that it has a version contract. If you ship on-device inference, move from commit-hash pins to v1.0.0 and track major versions for ABI breaks rather than chasing weekly builds.
- Verify local-inference claims on your hardware. Edge0's 145B numbers depend on 23 GiB of VRAM plus 38 GiB of host RAM on a 24 GB card and on expert residency; measure your own decode rate before designing a product around it, and read the custom licence.
- Treat self-hosted agents as a security-boundary decision, not a privacy checkbox. Anthropic keeps model calls leaving your network with full tool inputs and outputs, so custody of history and credentials is what you gain, along with the duty to patch the server yourself.
- If you consume Chinese open weights, add a provenance note to your model inventory. Weight forensics now gives a reproducible method for inferring training hardware, and the K3 and DeepSeek V4 lines are the first models where that inference is public.
Model releases & progress
Shipped in this window: Gemini 4 Hydrogen (Google, generally available in the Gemini API, AI Studio, Gemini app, Search, Gemini for business and Antigravity; $1.50/$7.50 per million tokens) and GPT-6.1 Terra (OpenAI, generally available in ChatGPT Work, Codex and the API; $1/$5, 1.05M context). Anthropic shipped no new checkpoint in the window: self-hosted Claude Managed Agents entered private beta on November 12, its newest models remain Claude Fable 5.2 and Mythos 5.2 from November 9, and Claude Opus 5.5 from September 22 is still the reference point in third-party comparisons. Open weights moved twice: Edge0 published 4-bit 145B and 120B agent models under a custom community licence, and IBM documented its LLMON previews for Granite 4.1 3B and 8B. Tooling crossed a threshold with llama.cpp and ggml v1.0.0. Reported but not first-party confirmed: a weight-forensics analysis finding Blackwell initialization traces in six Chinese open-weight models.
Bottom line
The frontier did not move today; the price of reaching it did. Google and OpenAI both shipped fast execution tiers with roughly frontier-level agent scores at $1–$1.50 per million input tokens, llama.cpp turned local inference into a versioned dependency, and Nvidia's leadership shock is the one open variable that could reprice everything downstream. Meta made no first-party model announcement in the window: its AI blog's most recent post is still Muse Spark 1.1 from July.
As of
November 13, 2026, 15:00 UTC. Research window: November 11–13, 2026. Status terms follow first-party wording: "available now" and "generally available" for shipped APIs and weights, "private beta" where access is gated, "reported" for claims not confirmed by the company. Benchmark figures are attributed as vendor-run unless stated otherwise, and scores from different harness versions are not directly comparable.