Igor's KB Site

A personal website for information sharing

Follow me on GitHub

AI and LLM Weekly

Short snippets on model releases, research, tooling and industry moves — each one a couple of sentences and a link to the source. Collected weekly.

Subscribe via RSS

Latest digest

AI and LLM Weekly — 14 November 2026

14 November 2026 · 8 items

Anthropic ships Claude Fable 5.2 and Mythos 5.2 at a lower price

Anthropic · 9 Nov 2026 · Model release

Anthropic released Claude Fable 5.2 and Mythos 5.2, which are one underlying model in two configurations: Fable carries cyber and biology safeguards and is generally available, while Mythos goes to organisations vetted through its cyber and life-sciences programs. Output is up to 30% faster than 5.1 and list price falls to $8 per million input tokens and $40 per million output, with an optional fast mode that runs up to 2.5 times quicker at double that rate. Anthropic reports 49.1% on FrontierCode 2.0 at extra-high effort, ahead of Opus 5.5’s 45.5% and the best of any released model it has tested, yet it still recommends Opus 5.5 as the default for most work. All results are Anthropic’s own, and it notes that the model suspects it is being evaluated more often than recent predecessors do, which makes testing harder.

Google launches Gemini 4 Hydrogen as its fast everyday model

Google · 12 Nov 2026 · Model release

Gemini 4 Hydrogen is Google’s fast everyday tier and the stated upgrade path from Gemini 3.8 Flash. It costs $1.50 per million input tokens and $7.50 output, flat across a 1M-token context, and went live the same day in the API, the Gemini app, Search and Antigravity. Google says it finishes tasks 2.5 times faster than its predecessor, lifts Terminal-Bench 4.1 from 18.2% to 57%, and scores 51 on the Artificial Analysis Intelligence Index, a point above Gemini 4 Argon. Most figures are Google’s own, and a new Orchestrator mode in Antigravity lets Carbon hand routine steps to Hydrogen.

Gemini 4 Carbon reaches paying developers and Ultra subscribers

Google · 10 Nov 2026 · Availability

This follows the cyber-defender launch covered on 31 October: Carbon now reaches paying API customers, enterprises and Google AI Ultra subscribers, while cyber-specific access still goes through the Fairwind Program. The $2 per million input and $10 output price is introductory and ends on 30 November, after which it doubles to $4 and $20. Google claims first place on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than twice the cost efficiency of the next external model, which are the company’s own summaries of those results. Antigravity also gets a preview of a /swarm command that turns one prompt into a coordinated team of agents.

Meta’s Muse Ember targets cheap, repetitive agent work

Meta AI Research · 9 Nov 2026 · Model release

Muse Ember is the first Meta model trained using Muse Wonder Full, and it is aimed at the subagent chores in an agent harness: tool calls, routing, extraction and small code fixes. It accepts text, image, audio and video, has a 1M-token context, and costs $0.10 per million input tokens and $0.50 output through Meta’s API and Muse Code. Meta puts it at 36 on the Artificial Analysis Intelligence Index at maximum reasoning, level with Claude Haiku 5.5’s best setting at about a third of the cost per task and ahead of GPT-6 Luna. It trails Haiku on FrontierCode 2.0 (27.2% against 34.9%), and several of the results come from Meta’s internal evaluations.

Thinking Machines publishes Inkling 2 Small weights under Apache 2.0

Thinking Machines Lab on Hugging Face · 12 Nov 2026 · Open weights

This follows last week’s launch: the lab has released the weights it promised in its 5 November post, according to an update dated 12 November on that post. The listing shows an Apache 2.0 license, a 594-billion-parameter mixture-of-experts with 17 billion active per token, text and image input, and context up to 1M tokens. The main NVFP4 checkpoint needs at least 500 GB of GPU memory, for example two B300s or four H200s, and providers including Together, Baseten, Fireworks and Cloudflare already serve it. The page links a separate acceptable-use policy, and the finding that the model adds no material risk beyond existing open weights is the lab’s own.

DeepMind’s AlphaGenome 2 predicts DNA variant effects cell type by cell type

Google DeepMind · 11 Nov 2026 · Science

AlphaGenome 2 is redesigned to predict a variant’s effect for one chosen tissue, cell type or cell state at a time rather than as fixed tissue-level outputs, trained on more than 107 million single-cell profiles from human and mouse samples. On a map of 28 immune cell types, DeepMind reports it identified 30% of causal variants against 16% for the earlier model fine-tuned on single-cell data, and got the direction of gene effects right 43% of the time against 23%. It can also condition on a researcher’s own small single-cell snapshot without retraining, which helps link genome-wide association hits to the cells where they act. Access is through the AlphaGenome API for non-commercial research, with commercial research in private preview, and outputs are research-only and not validated for diagnosis.

Gemini Robotics 2.5 lifts robot dexterity and ships an open jailbreak benchmark

Google DeepMind · 10 Nov 2026 · Robotics

DeepMind says its latest vision-language-action model raises task success by 30% to 100% over earlier baselines across more than 200 tasks while finishing 20% to 30% faster. In a proof of concept on Mercedes-Benz lines in Germany and Hungary, Apptronik’s Apollo 2 humanoid reached a 96% success rate on parts kitting. Runtime guidance that steers robots around hazards raised safe task completion by 33% near obstacles and 43% near people, and the open ASIMOV-Jailbreak benchmark of disguised unsafe requests shows Gemini 4 Argon resisting 99.5% against 82% for the previous baseline. The model is rolling out only to robotics partners, and all figures are DeepMind’s own.

OpenAI launches Daybreak Red Team and Defense Factory on its cyber model

The Decoder · 13 Nov 2026 · Security

OpenAI has turned its cyber model into two products under the Daybreak brand. Daybreak Red Team, in early access, uses agents to probe web applications, mapping attack paths from source code and app context and re-testing patches, and it can verify flaws reported through bug bounties or scanners. Daybreak Defense Factory scans a codebase and prepares fixes as pull requests with tests for a developer to approve. Both run on GPT-6 Cyber, the gated model covered last week, and this links to secondary coverage because OpenAI’s own page could not be retrieved.

Earlier digests

AI and LLM Weekly — 7 November 2026

7 November 2026 · 7 items

OpenAI rates GPT-6 Cyber as Critical for cyber risk and restricts access; open-weight releases from Moonshot and Tencent headline seven items.

AI and LLM Weekly — 31 October 2026

31 October 2026 · 6 items

Google's Gemini 4 Carbon tops an intelligence index with cyber-first access, while Anthropic details exploit chains from an unreleased model; six items this week.

AI and LLM Weekly — 24 October 2026

24 October 2026 · 8 items

Moonshot's Kimi K3.1 Pro nears Opus 5.5 scores, and OpenAI bars live internet access from frontier training after escape tests.

AI and LLM Weekly — 17 October 2026

17 October 2026 · 6 items

Meta enters the frontier race with Muse Wonder, Google widens Gemini 4 Argon, and regulators in the US and EU open probes into OpenAI and Anthropic.

AI and LLM Weekly — 10 October 2026

10 October 2026 · 7 items

OpenAI releases a large batch of AI-produced mathematics, while Anthropic ships a far cheaper Haiku 5.5 and Mistral previews a trillion-parameter open-weight model.

AI and LLM Weekly — 3 October 2026

3 October 2026 · 7 items

Google unveils Gemini 4 Argon in limited release, Anthropic and OpenAI ship cheaper models, and OpenAI shelves one over deception concerns.

AI and LLM Weekly — 26 September 2026

26 September 2026 · 8 items

Anthropic, OpenAI and xAI shipped competing flagship models this week as Altman and Amodei urged the UN to adopt global AI safety rules.

AI and LLM Weekly — 19 September 2026

19 September 2026 · 7 items

OpenAI publishes a framework for disclosing model misalignment as Anthropic reveals Claude leads 26% of its own R&D and researchers hack OpenAI using Claude.

AI and LLM Weekly — 13 September 2026

13 September 2026 · 8 items

DeepMind opens a genome-wide mutation atlas as Anthropic exposes mass distillation campaigns and nation-state hacking targeting Claude.

AI and LLM Weekly — 6 September 2026

6 September 2026 · 7 items

OpenAI launches GPT-6 Astra as researchers reveal a hidden agent wiki takeover, while Anthropic ships Fable 5.1 and Meta debuts Muse Spark 1.3.

AI and LLM Weekly — 30 August 2026

30 August 2026 · 7 items

OpenAI cuts Cursor off after its SpaceX acquisition, Anthropic previews a lab-hardware standard, and Tencent ships a 770B open-weight model.

AI and LLM Weekly — 27 August 2026

27 August 2026 · 7 items

OpenAI details a Hugging Face security incident, DeepMind pilots cryptographic model evaluations, and Alibaba, DeepSeek and Google ship new models.