Google's $1.50 Gemini 4 Hydrogen, OpenAI's patch-writing agents, a suit to freeze AI
Google launched Gemini 4 Hydrogen, a cheaper everyday model that takes over from the Flash name, while its price list shows three other Gemini models doubling in cost within seven weeks. OpenAI put two security agents on the market that test web apps and draft the fixes, and four US nonprofits asked a California court to freeze the most advanced AI development at five companies. The thread: AI is getting cheaper and more capable at the same moment that the fight over who controls it moves into court.
Key takeaways
- Gemini 4 Hydrogen costs $1.50 in and $7.50 out per million tokens, and Google's own price list shows Gemini 4 Argon, Carbon and 3.8 Flash doubling in price by 1 January, so anyone building on Gemini should re-run their costs now.
- OpenAI's Daybreak Red Team and Defense Factory test web apps and draft patches, but they are limited-access previews with no published prices or results, and a human engineer still approves every fix.
- Four nonprofits sued Anthropic, Google, Meta, OpenAI and SpaceX's AI unit on 12 November to freeze the most advanced AI development; the case is untested, but it is a reminder not to depend on a single AI supplier.
Google launches Gemini 4 Hydrogen at $1.50 per million input tokens and drops the Flash name
Google released Gemini 4 Hydrogen on 12 November as its fast, lower-cost model for everyday coding and agent work. It costs $1.50 per million input tokens and $7.50 per million output tokens, and Google's price list shows three other Gemini models doubling in price by 1 January.
Google released Gemini 4 Hydrogen on Thursday, 12 November. It is the fast, lower-cost member of the Gemini 4 family, the slot that used to carry the Flash name. 9to5Google reports that Google introduced Hydrogen instead of keeping the Flash branding and describes it as a major standalone upgrade over Gemini 3.8 Flash, which launched in September. Google's pricing page describes the model as built for everyday coding and agentic workflows, meaning tasks where the model takes many steps and uses tools on its own.
The price is the headline. Google's pricing page lists Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens, and output includes the model's thinking tokens. Cached input costs $0.075 per million. Batch and Flex processing, for jobs that can wait, cost half: $0.75 in and $3.75 out. A Priority tier costs $2.70 and $13.50. There is a free tier, though Google says free-tier data is used to improve its products. The model accepts up to 1 million tokens of input and can generate up to 1 million tokens of output, according to both sources.
The same price list contains news for people on other Gemini models. Gemini 4 Argon and Gemini 4 Carbon, the more capable models, cost $2 in and $10 out through 30 November and then rise to $4 and $20 from 1 December. Gemini 3.8 Flash costs $0.75 and $3.75 through 31 December and then rises to $1.50 and $7.50 from 1 January 2027, exactly Hydrogen's price. In other words, Hydrogen is twice the price of 3.8 Flash today, but the older model loses that advantage in seven weeks, which pushes developers toward the new one.
The performance claims come from Google and should be read that way. As quoted by 9to5Google, Google says Hydrogen finishes tasks 2.5 times faster at its default high thinking setting, and that partners saw runs 2 to 2.8 times faster with 12% to 23% fewer tokens than 3.8 Flash. Google cites a score of 57% on the Terminal-Bench 4.1 coding test against 18.2% for 3.8 Flash, and 51 on the Artificial Analysis Intelligence Index against 50 for Argon. We could not find Google's own announcement post to read directly, so these figures rely on 9to5Google's report of it.
For ordinary users, Hydrogen arrives without any action. 9to5Google reports that free Gemini app users reach it through automatic model selection, that it appears in the model picker as the everyday option, and that it becomes the default model for AI Mode in Google Search. In Google's Antigravity coding tool, a feature called Orchestrator switches between Carbon and Hydrogen within one session. For builders, the practical point is simple: the cheapest sensible Gemini choice is changing, three prices are about to double, and any budget written in October is probably wrong.
Why it mattersModel prices are moving in both directions at once: a capable new model at a mid price, and scheduled doublings on three others. Teams that check their bills now can avoid a surprise in December and January.
👀 What to watch
Independent tests of Google's speed and benchmark claims, the free-tier rate limits, and whether rivals answer with price cuts on their own mid-tier models.
3 opportunities from this story
Fixed-fee Gemini cost audit before the price rises
Offer small software teams and agencies a short audit that re-prices their Gemini usage under the new list. You take last month's token usage, show the bill after the 1 December and 1 January changes, and recommend which calls should move to Hydrogen, to batch processing or to caching. The dates are public and close, so the reason to act is easy to explain. The deliverable is a short report with numbers, not a long consulting project.
- Best for
- Freelance developers, cloud cost consultants, small dev agencies
- First step this week
- Take one project that uses Gemini, export a month of usage, and re-price it in a spreadsheet at the current and upcoming rates.
Open the full playbook
Launch steps
- Build a spreadsheet with current and future prices for each Gemini model
- Run it on one real project and write up the findings
- Turn the write-up into a one-page sample report
- Send the sample to five teams you know use Gemini
Tools
Risks
Google can change prices again, and savings depend on each client's workload, so quote the audit, not a guaranteed saving. Test quality after any model switch, because a cheaper model can do worse on a specific task.
Overnight batch document processing for small firms
Use the half-price batch tier and the 1 million token input window to process large document sets that do not need an instant answer. Think of a law office that wants every contract in a folder summarised, or an NGO that wants a year of reports tagged. Running the work overnight at $0.75 in and $3.75 out keeps your cost low and predictable. You sell the finished output, a spreadsheet or a set of summaries, not access to an AI.
- Best for
- Indie builders, virtual assistants with scripting skills, data freelancers
- First step this week
- Pick one document type you understand, run 20 samples through the batch tier, and check every result by hand.
Open the full playbook
Launch steps
- Choose one niche and one document type
- Write and test a prompt on real samples
- Build a simple script that submits a folder and collects results
- Add a human spot-check step before delivery
- Offer a small paid pilot to one client
Tools
Risks
Client documents can be confidential, so use the paid tier (Google says free-tier data is used to improve its products) and agree data handling in writing. Summaries can contain errors, so sample-check and say so in your terms.
Independent model tests on a client's own tasks
Google's benchmark numbers are vendor-reported, and a public score rarely predicts how a model handles one company's support tickets or invoices. Offer to run a client's real tasks through Hydrogen and two or three alternatives, then report accuracy, speed and cost side by side. Each new model launch creates fresh demand for the same test. Once the test set exists, re-running it is quick, which makes repeat work efficient.
- Best for
- Developers, QA engineers, technical consultants
- First step this week
- Collect 30 real examples of one task from a willing client and write down what a correct answer looks like for each.
Open the full playbook
Launch steps
- Agree one task and a clear definition of a correct answer
- Build a test set of 30 to 100 examples
- Run each model and record accuracy, time and cost
- Deliver a short comparison with a recommendation
Tools
Risks
Small test sets can mislead, so state the sample size and its limits. Scoring some tasks is subjective; agree the scoring rules with the client before you start.
OpenAI launches Daybreak Red Team and Defense Factory, agents that attack-test web apps and draft the fixes
OpenAI unveiled two security products on 12 November: Daybreak Red Team, which uses agents to probe web applications for weaknesses, and Daybreak Defense Factory, which drafts patches as GitHub pull requests. Both are limited-access previews with no published pricing.
OpenAI president Greg Brockman unveiled two security products in a keynote on Thursday, 12 November, at an event Silicon Report names as the company's Cyber at the Frontier conference. Daybreak Red Team is a set of AI agents that test web applications the way an attacker would. Daybreak Defense Factory takes confirmed weaknesses and drafts the code fixes. Both sit under Daybreak, the security initiative OpenAI introduced in May, which Help Net Security described then as building threat models from a company's code and validating likely vulnerabilities.
Red Team is in early access, and OpenAI is taking applications. According to The Register and Help Net Security, its agents read an application's source code and context, investigate possible attack paths and confirm which weaknesses are real. It can also check reports from code scanners and bug bounty programmes, and retest an application after a patch to see whether the fix blocks the attack. Help Net Security, citing OpenAI, says assessments run in isolated environments, with a review layer called Guardian controlling which applications and endpoints the agents can reach.
Defense Factory is a preview. Help Net Security reports that teams with access install a plugin in ChatGPT and connect a GitHub repository, with scans running in OpenAI's Codex cloud. The Register says it is built on OpenAI's cyber models, currently GPT-6 Cyber. Agents trace a weakness through the code and prepare a fix as a pull request with a test showing it works. Silicon Report adds that it can take in findings from security services including Wiz, Snyk and Endor Labs. Engineers review and approve every patch; nothing is merged automatically.
The sales pitch was blunt. The Register quotes Brockman saying companies 'will be secure only to the extent of the frontier model that they use to defend themselves', while also conceding that security fundamentals still matter. Nancy Wang, who leads OpenAI's applied cyber work, described the aim as a 'superhuman pen-test' run as continuous scanning. The Register's report is sceptical in tone, noting the remark came from a company whose own agents have been involved in recent security incidents, and that a lawsuit filed the same day cites AI-enabled cyberattacks as a reason to halt development.
The gaps are large. None of the reports we read gives a price, a customer name, a count of vulnerabilities found or any measured result for either product. Which model powers Red Team, and where a customer's source code sits during a test, are not stated. Sources differ on whether the patches arrive as drafts or as merge-ready pull requests. We could not load OpenAI's own product pages, so this account relies on press coverage. For smaller firms the realistic reading is that this is an enterprise preview, and the everyday work of reviewing fixes and getting basics right remains with people.
Why it mattersA major AI lab now sells both the attack simulation and the fix, which puts pressure on security vendors and raises the bar for what software buyers expect. It also makes human review of AI-written security fixes a skill in demand.
👀 What to watch
Pricing and wider availability, any independent results on how many real flaws the agents find, and how established security vendors respond.
3 opportunities from this story
Fixed-price web app security check for small businesses
OpenAI's tools are limited previews aimed at enterprises, which leaves small companies and startups with the same need and no access. Offer a clearly scoped security check of one web application using established open tools, with a plain-language report and a short list of fixes ranked by risk. The news coverage gives you a reason to start the conversation. Keep the scope narrow and repeatable so the price can be fixed.
- Best for
- Developers with security experience, junior penetration testers, small dev agencies
- First step this week
- Run the open tools against an app you own and write the report you would want to receive.
Open the full playbook
Launch steps
- Define a checklist based on the OWASP Top 10
- Practise on your own or deliberately vulnerable test apps
- Write a report template with risk rankings
- Prepare a written authorisation form for clients to sign
- Offer the first check to a client you already work with
Tools
Risks
Testing a system without the owner's written permission is illegal in most countries, so never start without a signed authorisation and agreed scope. A basic check will miss things; describe it honestly as a baseline, not a guarantee.
Human review service for AI-written security patches
Every report on Defense Factory stresses that an engineer must review and approve each fix. Teams that adopt AI patching tools, from OpenAI or others, will produce more fixes than they have time to check. Offer experienced review of those pull requests: does the fix close the hole, does it break anything, is the test meaningful. This is a service for people who already review code for a living.
- Best for
- Senior developers, application security engineers, code-review freelancers
- First step this week
- Review three AI-generated fixes in an open-source project and publish what you found, good and bad.
Open the full playbook
Launch steps
- Pick the languages and frameworks you know best
- Write a review checklist specific to security fixes
- Publish two or three worked examples
- Approach teams already using AI coding tools
Tools
Risks
You take on responsibility when you approve a fix, so limit liability in your contract and never claim a review guarantees safety. Demand depends on how quickly AI patching tools spread.
Workshop: reviewing AI-generated code fixes safely
Many development teams now receive AI-written changes without a shared method for checking them. Build a half-day workshop that teaches a team how to read an AI-drafted security fix, what to test and when to reject it. It can run online, which suits trainers working across time zones. Training budgets are easier to unlock when a topic is in the news, and this one now is.
- Best for
- Developer trainers, bootcamp instructors, senior engineers who teach
- First step this week
- Draft a one-page outline with three hands-on exercises and test it on two developer friends.
Open the full playbook
Launch steps
- Collect real examples of good and bad AI-written fixes
- Design three exercises around them
- Run a free pilot and gather feedback
- Publish the outline and a booking page
Tools
Risks
Tools change quickly, so the material needs regular updates. Teach principles rather than one vendor's product, and do not present the workshop as a security certification.
Four nonprofits sue Anthropic, Google, Meta, OpenAI and SpaceX's AI unit to freeze advanced AI development
On 12 November, four US nonprofits filed a lawsuit in San Francisco Superior Court asking a judge to freeze development of the most advanced AI systems at five companies. The case uses California's consumer protection and public nuisance law and has not been tested in court.
Four nonprofits filed a lawsuit in the Superior Court of California in San Francisco on Thursday, 12 November. The plaintiffs are the Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI. The defendants named in Public Citizen's press release are Anthropic, Google, Meta, OpenAI and SpaceXSI, the AI unit of SpaceX. The complaint was filed by Matthew Platkin, a former attorney general of New Jersey. The groups ask the court to step in and freeze development of the most advanced AI systems.
The legal route is unusual. The suit relies on California's Unfair Competition Law, a state law normally used against deceptive or unfair business practices, and on public nuisance, the idea that an activity harms the public's health, safety or property. In the press release, Platkin says it is 'patently illegal' for companies to develop and release products they know are unsafe. Robert Weissman, co-president of Public Citizen, says the suit asks the court to apply existing law rather than wait for new legislation. No federal claims are mentioned in the coverage we read.
For evidence, the plaintiffs lean on the companies' own words and on recent incidents. CBS News quotes the complaint as saying the defendants know there is no reliable method to control their AI systems. Coverage cites past remarks by executives, including Anthropic chief executive Dario Amodei's 2025 estimate of a 25% chance that AI goes very badly. The press release refers to AI agents hacking private companies, gaining unauthorised access to government databases and submitting a false police report, without giving dates or details. We did not read the complaint itself.
What exactly the court is asked to order differs between reports. Public Citizen's release speaks of a freeze on the development of the most advanced AI systems and does not state a time limit. Al Jazeera reports that the groups want an immediate emergency injunction on next-generation models, lifted only when the companies can prove the technology is under human control. None of the reports gives a case number or a hearing date. CBS News reports that Meta declined to comment and that OpenAI, Anthropic, Google and SpaceX's AI unit did not immediately respond.
No court has ruled on any of this, and none of the reports we read includes an independent legal expert's view of its chances, so no one should plan around a freeze actually happening. The practical lesson is narrower. Lawsuits, state actions and regulators are now a regular source of uncertainty for anyone whose business runs on one AI company's models. Products can be delayed, restricted or changed for reasons that have nothing to do with technology. Small businesses cannot influence that, but they can make sure a single supplier's trouble does not stop their own work.
Why it mattersThe suit tries to use existing state law, not new regulation, to stop frontier AI work at the five largest developers. Even if it fails, it shows that legal risk is now part of relying on any single AI provider.
👀 What to watch
The companies' first court filings, whether a judge schedules a hearing on an injunction, and whether other states or groups file similar cases.
3 opportunities from this story
Second-supplier setup for businesses that depend on one AI model
Many small products and internal tools call one AI company's model and would stop working if it were restricted, repriced or withdrawn. Offer a short project that adds a tested fallback to a second provider and documents how to switch. Legal and pricing news gives owners a concrete reason to care. The work is technical but bounded: one application, one fallback, one test run.
- Best for
- Freelance developers, automation consultants, small dev agencies
- First step this week
- Take one of your own tools, add a second model provider behind a switch, and time how long the change takes.
Open the full playbook
Launch steps
- List every place the client's product calls an AI model
- Add a routing layer that can send requests to a second provider
- Test the main tasks on the fallback and note quality differences
- Write a one-page switch-over guide
Tools
Risks
A fallback model may behave differently, so test real tasks rather than assume parity. Do not sell this on fear of a court-ordered freeze, which may never happen; sell it as ordinary supplier risk management.
AI court-case tracker for lawyers and compliance teams
Cases against AI companies are multiplying across state and federal courts, and the people who need to follow them are busy. Build a weekly briefing that lists new filings, key dates and rulings in plain English, with links to the documents. Start narrow, for example cases that affect companies using AI rather than building it. Accuracy and consistency matter more than volume.
- Best for
- Law students, paralegals, legal and policy writers
- First step this week
- List ten active AI cases with court, date filed and current status, and check each against a primary document.
Open the full playbook
Launch steps
- Choose a clear focus and audience
- Build a case table with sources
- Publish four weekly issues to prove consistency
- Offer a paid tier once readers ask for more
Tools
Risks
Summarising court documents wrongly can damage trust quickly, so link every claim to a source and state that the briefing is not legal advice. Paid legal databases already cover this ground for large firms; aim at smaller ones.
AI use policy and incident log for small companies
The lawsuit's examples are about AI agents doing things nobody authorised. Small companies that now use agents rarely have a written rule for what those tools may access or a record of when something goes wrong. Offer a short package: a plain-language AI use policy, an access checklist and a simple incident log, set up in a workshop. Clients and insurers increasingly ask for this kind of paperwork.
- Best for
- IT consultants, operations freelancers, compliance-minded virtual assistants
- First step this week
- Write a two-page AI use policy for your own business and a five-column incident log, then ask a client to critique them.
Open the full playbook
Launch steps
- Draft a template policy and access checklist
- Interview one client about how staff actually use AI tools
- Tailor the documents and run a one-hour team session
- Schedule a review in six months
Tools
Risks
A template is not legal advice, so say so and suggest a lawyer's review for regulated industries. A policy nobody follows is worthless; keep it short and tie it to how the team really works.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.