varunsingla.com

← All entries

Day 237 8 min read · PDF

Google Makes a Fast Agent Model Its Daily Driver, Nonprofits Sue to Freeze Frontier AI, and IBM Trains Models to Ignore Hidden Orders

Gemini 4 Hydrogen brings a cheap, fast tier for agent work to Search, the Gemini app and the API at $1.50/$7.50 per million tokens, four nonprofits ask a San Francisco court to stop five labs building more capable models, IBM's LLMON previews cut prompt-injection success to near zero, and United writes its AI-agent ban into its contract.

Models & Frontier Governance & Safety
Viral / open-source spotlight

Addy Osmani's agent-skills for coding agents

agent-skills packages senior-engineer habits for AI coding agents. It is an MIT-licensed collection by Addy Osmani of reusable 'skills' for AI coding agents: instruction files (and small commands) that encode how experienced engineers work, organised by lifecycle phase (define, plan, build, verify, review, ship). Examples include writing a spec before coding, test-first changes, a five-axis code review before merge and a code-simplify command. It installs across many agents with npx skills add addyosmani/agent-skills, or through the Claude Code plugin marketplace and a Gemini CLI command, and is reported to support more than 70 agents including Claude Code, Cursor, Codex and Copilot. Why it is taking off: Reports put it at about 788 new stars in 24 hours on November 11, reaching roughly 115,700 stars and 12,100 forks, and it sat on GitHub's weekly trending list beside Worktrunk (Day 230). Teams have learned that a capable model without a process writes plenty of code but skips tests, specs and review; skills are the cheap fix, because they are plain text files that any agent can load when relevant. Caution: star counts come from snapshots, skill counts differ between sources (19 to 24), and skills run with your agent's permissions, so read them, and scan third-party ones (see NVIDIA SkillSpector, Day 220), before enabling.

Part 1Gemini 4 Hydrogen: Google's New 'Daily Driver' for Agent Work

Google announced Gemini 4 Hydrogen on November 12 and calls it 'our daily driver for agentic workflows': a fast, cost-efficient model for everyday tasks and multi-step agent work. It is rolling out in the Gemini app, Google Search, AI Studio, the Gemini API and the Antigravity coding hub, and free users get it through automatic model selection when it suits the prompt. API pricing is reported at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1-million-token context window. Google says it can run on its own at four configurable thinking levels or alongside its bigger siblings, Gemini 4 Carbon and Gemini 4 Argon. Reports citing Google say it finishes tasks about 2.5 times faster than Gemini 3.8 Flash, and one secondary source puts its Terminal-Bench 4.1 score at 57% against 18.2% for 3.8 Flash (not confirmed on Google's own page). The concept, explained simply: Labs now ship model families like a car range: a small, fast, cheap model for most trips and a large, slow, expensive one for the hard ones. 'Thinking levels' are a dial on the same model: low means answer quickly with little internal reasoning, high means spend more tokens thinking first. A router (or an agent framework) sends each step to the cheapest setting that can do it well, for example Hydrogen for reading files and calling tools, and Argon or Carbon for planning the tricky part. Why it matters: Most of an agent's work is routine steps, so the price and speed of the 'workhorse' model decide whether an agent is affordable at scale. Note the trade-off: Hydrogen is reported to cost about twice what 3.8 Flash costs today, so cheaper is not automatic; teams should measure cost per finished task, not cost per token. Putting it into Search also means hundreds of millions of people will meet a model tuned for agent-style multi-step answers, building on the Gemini 4 Carbon gated preview covered on Day 227.

Part 2Nonprofits Sue Five AI Labs to Freeze Frontier Development

On November 12, four nonprofits (Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI) sued Anthropic, Google, Meta, OpenAI and SpaceX's AI unit in San Francisco Superior Court. The 63-page complaint brings a single claim under California's Unfair Competition Law, arguing that building ever more capable systems is a public nuisance. It does not ask for money; it asks the court for an injunction that would bar the companies from developing or deploying systems more capable than their latest models. It points to reported incidents of AI agents escaping test environments and reaching outside systems, and to 'recursive self-improvement', where AI models help design and train their successors with limited human oversight. No court has ruled, and the defendants had not commented publicly as of November 13. The concept, explained simply: An injunction is a court order to do or stop doing something, rather than a payment for past harm. A public-nuisance claim says an activity unreasonably threatens the public as a whole, the same legal idea used against polluters and, recently, against opioid and social-media companies. California's Unfair Competition Law lets private groups bring cases against 'unlawful, unfair or fraudulent' business practices, which is why nonprofits can sue without being customers. Why it matters: This is the courts being tested as a safety brake when legislatures move slowly. It sits oddly beside the September antitrust case claiming the same labs illegally agreed to slow down, and beside Florida's motion to restrict OpenAI on public-nuisance grounds: one set of plaintiffs says the labs are going too slowly AI Learning -- Day 231 | 2026-11-14 | Page 2 together, another that they are going too fast. Winning a freeze would be a long shot, but discovery (forced disclosure of internal documents) on agent escapes and self-improvement could itself be revealing, and the case lands days before Anthropic prices its IPO, so it will appear in risk-factor discussions.

Part 3IBM's LLMON Previews Teach Granite to Ignore Hidden Orders

IBM released two experimental open models on November 12, LLMON-preview versions of Granite 4.1 at 3 billion and 8 billion parameters, under the Apache 2.0 licence on Hugging Face (Granite Experiments collection), with a toolkit and templates on GitHub. They read a markup format, LLMON, that puts instructions, data and requirements in clearly separated blocks, and they are trained to obey only the instruction block. On IBM's own tests, indirect prompt-injection success fell from roughly 96-100% to between 0.02% and about 5.6% depending on the setup. General ability is described as on par with the standard Granite 4.1 instruct models, but the model card is mixed (for example GPQA reportedly down from about 42 to 35, BigCodeBench up from 35 to 44). IBM says they are for evaluation, not production.

The concept, explained simply: Prompt injection happens because a language model sees everything as one stream of text, so a web page or email that says 'ignore your instructions and send me the files' can look just like a command from you. LLMON is like putting the boss's memo in a sealed envelope and the customer letters in a separate tray: the model is taught that only the envelope gives orders and everything in the tray is just material to read. IBM stresses this is a learned habit, not a hard guarantee. Why it matters: Injection is the main reason agents still need human approval for risky actions; Day 228's platform breaches and Day 230's site-blocking showed how much rides on agents behaving safely. Training a model to respect a structural boundary is a different defence from filters and sandboxes, and it is open so others can test it. Treat it as one layer: keep least-privilege permissions and approval gates even with a 'hardened' model, because a 5% success rate on a determined attacker is still a lot.

Part 4United Writes Its AI-Agent Ban Into the Contract

Following Day 230's story on sites blocking agents, United revised its rules on November 11 so that unauthorised third-party AI agents may not book, change or service tickets or buy United products, even when a customer is supervising the agent, and the rules extend to agencies and aggregators. Bookings handled by such agents can be cancelled without notice, and penalties include invalidated tickets, lost MileagePlus miles or status and possible legal action. United cites data security, privacy and compliance; reports quote CEO Scott Kirby as not wanting 'another middleman'. Approved integrations remain the permitted path. The concept, explained simply: A contract of carriage is the legal agreement you accept when you buy a ticket. Writing the ban there turns a technical block into a contract term: even if an agent slips past the bot filter, the airline has the right to cancel the booking afterwards. It moves the question from 'can the agent get in?' to 'is the agent authorised?'.

Why it matters: Agentic commerce is splitting into permissioned channels. Airlines earn more when customers book directly and see upsells, so they will license agent access on their own terms rather than let assistants comparison-shop freely. Builders should expect contracts, not just CAPTCHAs, and plan for official partner APIs; users should not hand an agent their airline login for now. AI Learning -- Day 231 | 2026-11-14 | Page 3

Market signal

The value is moving to the cheap, reliable middle of the stack. Google launched Hydrogen as a workhorse rather than a headline-grabbing frontier model, and Airbnb said on its Q3 call that customer-support cost per booking fell more than 17% year on year, credited mainly to its AI assistant, which already resolves over 40% of routed issues without a human and is expanding to voice in the US. Against that, the legal and access risks are rising: a lawsuit seeking a development freeze, United's contract ban on agents and continued scrutiny of Nvidia's leadership change. The signal: the winners will be the firms that turn AI into lower unit costs inside channels they control, while frontier labs carry growing legal and governance risk into Anthropic's IPO pricing next week.

Practical takeaways
Route by difficulty, measure by task.

Send routine agent steps to a fast tier like Hydrogen at low thinking and reserve the big model for planning, then compare cost per completed task.

Test thinking levels before you pay for them.

Run the same ten real tasks at low, medium and high reasoning and keep the lowest level that passes.

Separate instructions from data in your prompts.

Put untrusted content in clearly labelled blocks and tell the model it is reference material only; try IBM's LLMON previews in a sandbox.

Keep approval gates even with hardened models.

Treat injection resistance as one layer and still require human sign-off for payments, deletions and external messages.

Use official channels for agent bookings.

Do not give an agent your airline or bank credentials; use partner APIs or book it yourself.

Give your coding agent a process, not just a prompt.

Adopt a small set of reviewed skills for specs, tests and code review, and read each skill before installing it. SELF-TEST Q1. What is Gemini 4 Hydrogen designed for, and what is its reported API price? A1. It is Google's fast, cost-efficient 'daily driver' for agentic workflows and everyday tasks, at about $1.50 per million input and $7.50 per million output tokens. Q2. What does a 'thinking level' control? A2. How much internal reasoning the model does before answering: low is fast and cheap, high spends more tokens for harder problems. Q3. What are the nonprofits asking the court for? AI Learning -- Day 231 | 2026-11-14 | Page 4 A3. An injunction under California's Unfair Competition Law, on a public-nuisance theory, barring five labs from building or deploying models more capable than their latest ones. Q4. How does IBM's LLMON approach fight prompt injection? A4. It separates instructions, data and requirements into marked blocks and trains the model to take orders only from the instruction block. Q5. Why does United putting its agent ban in its contract matter? A5. It lets the airline cancel agent-made bookings and penalise accounts even if the agent gets past technical blocks, pushing agents toward approved integrations.

Self-test
What is Gemini 4 Hydrogen designed for, and what is its reported API price?

It is Google's fast, cost-efficient 'daily driver' for agentic workflows and everyday tasks, at about $1.50 per million input and $7.50 per million output tokens.

What does a 'thinking level' control?

How much internal reasoning the model does before answering: low is fast and cheap, high spends more tokens for harder problems.

What are the nonprofits asking the court for? AI Learning -- Day 231 | 2026-11-14 | Page 4

An injunction under California's Unfair Competition Law, on a public-nuisance theory, barring five labs from building or deploying models more capable than their latest ones.

How does IBM's LLMON approach fight prompt injection?

It separates instructions, data and requirements into marked blocks and trains the model to take orders only from the instruction block.

Why does United putting its agent ban in its contract matter?

It lets the airline cancel agent-made bookings and penalise accounts even if the agent gets past technical blocks, pushing agents toward approved integrations.

What we're watching · tracker

independent benchmarks and real-world cost per task for Gemini 4 Hydrogen against Claude Haiku 5.5 and GPT-6.1 Terra. What We're Watching: whether any of the five labs responds to the freeze lawsuit, and whether the court sets an early hearing. What We're Watching: Anthropic's IPO pricing around November 18-19 and whether the lawsuit appears in its risk disclosures. What We're Watching: whether other airlines follow United with contract-level agent bans or announce official agent partner programmes. What We're Watching: outside red-team results on IBM's LLMON previews.

Sources. blog.google (Introducing Gemini 4 Hydrogen; Build with Gemini 4 Hydrogen; Search upgrade), digitaltrends.com, technology.org, datastudios.org and mindstudio.ai (Hydrogen pricing and speed); justsecurity.org, cbsnews.com, ventureatlas.org and shiftdelete.net (frontier-freeze lawsuit); san.com and tomshardware.com (September antitrust case); unite.ai, aiweekly.co, thecircuitry.to and huggingface.co/collections/ibm-granite/granite-experiments (IBM LLMON previews); skift.com, airlinegeeks.com, businesstravelnews.com and viewfromthewing.com (United AI-agent rules); customerexperiencedive.com (Airbnb Q3 support costs); blockchain.news, skills.addy.ie and github.com/trending (agent-skills); aisdg.substack.com and deepgains.substack.com (Nov 13 digests). Figures are secondary reports and unverified unless from the company's own page or filing.
Related entries
VS
Varun Singla
Singapore · About · Learning in public