← Daily LLM news

Daily LLM news, Friday, 13 November 2026

LLM Daily Briefing: Friday 13 November 2026

Google logo and brand image
Image: Google (blog.google)

Good morning, here's today's briefing. The lead: Google releases Gemini 4 Hydrogen, a faster, cheaper model for everyday coding and agent work. Also today: Thinking Machines opens the weights of Inkling 2 Small, IBM trains small Granite models to tell instructions apart from data, and OpenAI shows how automated jailbreaks made GPT-6 harder to talk into unsafe answers.

LEAD: Google launches Gemini 4 Hydrogen for everyday coding and agents

Background: Earlier this week Google made Gemini 4 Carbon generally available as a frontier model for hard, long-horizon work. Teams still needed a faster, cheaper model for the everyday loop of writing code, running tests and calling tools.

What happened: On Thursday Google launched Gemini 4 Hydrogen, now rolling out across Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, Google Sheets, and paid Google AI Pro and Ultra plans in Google Antigravity. Free Gemini app users get it when automatic routing picks it. Hydrogen is priced flat across a 1 million token context window at $1.50 per million input tokens and $7.50 per million output tokens. Google says it finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking setting, scores 57% on Terminal-Bench 4.1 against Flash's 18.2%, and posts 51 on the Artificial Analysis Intelligence Index, just ahead of Gemini 4 Argon's 50, or 54 at a new "xhigh" thinking level. There are four thinking levels in all: low, medium, high and xhigh. Google also launched Orchestrator in Antigravity, which can switch between Carbon and Hydrogen in a single session. On Terminal-Bench it reached near-Carbon success rates while using 24% less quota and finishing 19% faster. Orchestrator (Carbon) and Orchestrator (Argon) are rolling out to Ultra subscribers, with Orchestrator (Argon) for paid Pro subscribers. Google also reports safety gains, including up to 81% less overreach in agent tasks and more than 2.3 times more resilience on the Gray Swan indirect prompt injection benchmark. Google, Antigravity, model card

Why it matters: Google is splitting the work: Carbon for hard judgement, Hydrogen as the cheap, fast workhorse, and a router that switches between them mid-task rather than guessing the whole job from the first prompt.

Thinking Machines releases open weights for Inkling 2 Small

Background: Mira Murati's Thinking Machines Lab introduced Inkling 2 Small on 5 November as a closed-weight model on its Tinker platform, and said it would open the weights under an Apache 2.0 licence after outside safety testing.

What happened: On Thursday the company said the weights are now on Hugging Face, also integrated with Unsloth, SGLang, vLLM and TokenSpeed, and available through APIs including Nvidia, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters and 17 billion active, image input, a 1 million token context window, and a score of 31 on the Artificial Analysis Intelligence Index. The main NVFP4 checkpoint needs at least 500GB of combined GPU memory, for example two Nvidia B300s or four H200s. Thinking Machines, Hugging Face

Why it matters: A capable open-weight model from one of the best-funded independent labs gives teams a serious alternative they can run, fine-tune and inspect, if they can afford the hardware.

IBM's Granite models use a new markup language to resist prompt injection

Background: Long prompts mix system instructions, documents, tool outputs and user text, often in JSON, a format designed for parsers rather than language models. Attackers hide extra instructions in retrieved documents or tool results, and smaller models are especially easy to fool.

What happened: On Thursday IBM released two experimental preview models, Granite 4.1 3B LLMON and 8B LLMON, under an Apache 2.0 licence, plus a toolkit for a markup language it calls LLMON, pronounced "lemon." Instructions, data and requirements are marked with reserved special tokens and unique IDs, and an execution block says which instruction to run on which data. On the OPI prompt-injection benchmark, IBM says the models avoid 47% more attacks than the best frontier model it tested, and 95% more than the best open-source model, while staying close to Granite 4.1 Instruct on general tests. IBM is clear this is a learned behaviour, not a guarantee, and the previews are for evaluation, not production. IBM on Hugging Face, 3B model

Why it matters: Most safety filters try to catch a bad instruction after it has already landed in the prompt. IBM is trying to stop the model treating untrusted data as a command in the first place.

OpenAI: automated jailbreaks made GPT-6 harder to talk into unsafe answers

Background: In domains such as cybersecurity and biology, the same information can help a defender or an attacker, so it is hard to write a precise rule for when a model should refuse. OpenAI already gates some of those capabilities behind trusted-access programmes.

What happened: In a 12 November alignment post, OpenAI researchers described training GPT-6 Astra and Sol on jailbreaks found by an automated attacker. The attacker started from a seed prompt and kept rewriting it until the defender produced an unsafe reply, then those "robustified" prompts were used for reinforcement learning. In one example, a request about spreading anthrax spores failed at first, then succeeded when it was recast as a maths problem about airflow through outdoor vents. On a fixed set of earlier jailbreaks, GPT-6 Astra answered 91.5% of cyber prompts safely, against 59% for GPT-5.6 Sol, and it stayed more robust against adaptive multi-turn attackers. OpenAI says helpfulness on legitimate requests also improved, though the figures are for the model without production safeguards. OpenAI Alignment

Why it matters: As models get more useful in dangerous domains, safety depends less on a static list of banned prompts and more on whether training can keep up with attackers that iterate.

← Thursday, 12 November 2026 Sunday, 15 November 2026 →

Get the weekly sift

The best new AI and agent releases, with one-line verdicts. Occasional picks, no spam.

Privacy policy: your email goes to our email provider, Resend. Unsubscribe any time.

News is editorial and never paid for. Sponsored content on ToolSift is always labelled.