Gemini 4 Hydrogen, Cursor's OpenAI deadline and an AI freeze suit

Thursday 12 November had three stories for anyone who builds with AI models. Google released Gemini 4 Hydrogen, a fast model for everyday coding and agent work at $1.50 and $7.50 per million tokens, with a free tier. The date OpenAI proposed for cutting Cursor off from its models arrived, and a day later Cursor's own pages show no change. And four nonprofits sued five AI companies in San Francisco, asking a court to stop development of the most advanced models.

Here is what is confirmed in each story, what is not, and the one habit that covers all three.

Gemini 4 Hydrogen: Google's fast model for everyday coding

Two days after Gemini 4 Carbon, Google filled in the tier below it. Hydrogen is generally available, the API model ID is gemini-4-hydrogen, and Google's launch post says it is built for "everyday coding and agentic workflows", with Carbon and Argon kept for work that needs "deep architectural judgment".

Where you can use it

• Gemini API and Google AI Studio, on every tier including the free one, and in Android Studio.
• Google Antigravity, for paid Google AI Pro and Ultra subscribers.
• The Gemini app and Google Search, for everyone.
• Cursor: its models page already lists Gemini 4 Hydrogen next to Carbon and Argon.

Google's post and developer docs do not mention Gemini CLI, so do not assume it is there yet.

What it costs

Prices are per million tokens from Google's pricing page, checked on 13 November 2026. The Claude row is from Anthropic's pricing page, checked on 11 November.

Model	Input	Output	Notes
Gemini 4 Hydrogen	$1.50	$7.50	Free tier available. Cached input $0.075. Batch and Flex $0.75 / $3.75
Gemini 3.8 Flash (until 31 Dec)	$0.75	$3.75	Rises to $1.50 / $7.50 on 1 January 2027
Gemini 4 Carbon (until 30 Nov)	$2	$10	$4 / $20 from 1 December
Claude Sonnet 5.5	$2	$10

Three things to read out of that table:

• It is not cheaper than the model it replaces, yet. Hydrogen costs twice what Gemini 3.8 Flash costs today. Flash rises to the same price on 1 January, so from the new year Google's fast tier is $1.50 and $7.50 whichever one you pick.
• The free tier is a trade. Free-tier content is marked as used to improve Google's products. Paid-tier content is not. Do not point the free tier at a client's repository.
• Thinking tokens are billed as output. Hydrogen has four thinking levels: low, medium, high (the default) and a new xhigh. A request with 100,000 input tokens and 5,000 output tokens costs about $0.19 at the paid rate, and higher thinking levels raise the output side.

What Google claims

All of these are Google's own numbers from its launch post, and the post names no competitor models:

• Artificial Analysis Intelligence Index: 51 at the default high setting, one point above Gemini 4 Argon's 50, and 54 at xhigh.
• Terminal-Bench 4.1: 57%, against 18.2% for what the post's context indicates is Gemini 3.8 Flash.
• Speed: tasks finish 2.5 times faster than on 3.8 Flash at the default setting. Partner tests on SQL and data-heavy work ran 2 to 2.8 times faster and used 12% to 23% fewer tokens.

The number that got the headlines, "19% faster", is narrower than it sounds. It describes Orchestrator in Antigravity, a setup where Carbon coordinates and Hydrogen runs the worker and verifier steps. On Terminal-Bench 4.0, Google says that pairing came close to Carbon's success rate while using 24% less quota and finishing 19% faster. It is a claim about mixing models, not about Hydrogen alone.

That is the practical takeaway too. The pattern from the Carbon and Fable launches holds: an expensive model plans and reviews, a cheap fast one does the volume. If you already route work that way, Hydrogen is a new candidate for the cheap seat. I have not run it on a real project yet, so treat everything above as vendor numbers.

Cursor and OpenAI: the deadline passed, and nothing is confirmed

On 28 August, two weeks after SpaceX completed its acquisition of Cursor, OpenAI said it had told SpaceX it intends "to wind down our contract providing OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026." Its stated reason, as CNBC quoted it: "we cannot be confident that SpaceX will use our technology within our terms of service." CNBC also reported that OpenAI would not give Cursor future models during the wind-down.

Cursor's chief executive, Michael Truell, replied the same night that "OpenAI models serve about 5% of Cursor user traffic, and we're speaking with the OpenAI team to resolve this." Anthropic's Tom Brown said it would "continue to increase compute to support Claude models in Cursor."

That date was yesterday. Here is what I could and could not verify on 13 November:

• Cursor's models page still lists GPT models, with no removal or deprecation notice.
• Cursor has announced nothing. Its changelog, blog and forum announcements have no post about OpenAI models in November.
• Neither company has said in public that the date was confirmed, moved or dropped, as far as I can find. OpenAI's own pages block automated readers, so I could not check its latest wording directly.

The word that matters is "proposed". OpenAI never called 12 November final, and the two companies were still talking in August. The honest status is that GPT access in Cursor could end with little warning, or not at all.

What to do if you use GPT models in Cursor

1. Find everything that names a GPT model. Your saved model choice, team defaults, automations, and any script that calls the Cursor CLI with a model flag. Those are what break on the day.
2. Pick a fallback and test it now, on your own repository. Cursor bills Claude, GPT and Gemini the same way, at API list price from prepaid credit on Premium and above, so switching provider does not change how you pay. The Cursor profile covers the plans.
3. If you need GPT models specifically, plan for a tool OpenAI controls. OpenAI's own Codex is the one place its models cannot be withdrawn from by a contract dispute. Check OpenAI's current help pages for what an API key of your own still covers inside Cursor before you rely on it.
4. Keep your instructions portable. One AGENTS.md that any agent reads means a model or tool switch costs you an afternoon, not a week. The Cursor vs Claude Code post shows the setup.

The lawsuit asking a court to freeze frontier AI

Also on 12 November, four nonprofits sued OpenAI, Anthropic, Google, Meta and SpaceX's AI division in San Francisco Superior Court. The plaintiffs are the Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI. The complaint was filed by Matthew Platkin, a former attorney general of New Jersey.

Reuters describes the suit as "seeking to block the further development of the most advanced AI models". The plaintiffs argue that California's Unfair Competition Law gives a court grounds to order that, and that the companies' practices create a public nuisance. Public Citizen's co-president, Robert Weissman, said the suit asks the court to "freeze further AI development". The complaint cites incidents in which, in Reuters' words, models "escaped testing environments and autonomously attacked government websites."

Meta declined to comment to Reuters, and the other four had not responded when its story ran.

This is not the only case of its kind. Reuters notes that Florida's attorney general sued OpenAI in June and asked a judge in September to bar it from developing new models without outside oversight, a motion that is still pending, and that eight states sued OpenAI last month.

For a working developer, three points:

• Nothing changes today. There is no ruling. The coverage I read mentions no hearing date and no request for an emergency order.
• The target is future development. The reporting describes a bar on developing the most advanced models, not on access to models already released.
• It is still a new kind of risk. Until now the model under your tools changed because of price or product decisions. Courts are a third way, and you have no control over the timing.

One habit covers all three

A new model at a new price, a supplier pulling out of an editor, a lawsuit aimed at the labs themselves. In each case the model underneath your workflow turned out to be less fixed than it looked.

The defence is the same one from last week's post on multi-agent coding: keep the parts you own portable. Agent instructions in plain files, isolation and review in plain Git, and at least one tested fallback model from a different provider. Then a week like this one is a settings change.

The short version

• Gemini 4 Hydrogen is out: gemini-4-hydrogen, $1.50 and $7.50 per million tokens, free tier included, aimed at everyday coding. It costs double Gemini 3.8 Flash until Flash's price rises on 1 January. The "19% faster" figure is about Carbon and Hydrogen working together in Antigravity.
• Cursor and OpenAI: the proposed 12 November cutoff has passed with no public confirmation either way, and Cursor still lists GPT models. If your setup depends on them, test a fallback this week.
• The freeze lawsuit asks a California court to stop development of the most advanced models at five companies. No ruling, no dates, nothing changes for released models today.

Frequently asked questions

How much does Gemini 4 Hydrogen cost?

On the paid tier of the Gemini API, $1.50 per million input tokens and $7.50 per million output tokens, with cached input at $0.075. Batch and Flex are half that. There is also a free tier, where Google may use your content to improve its products. Checked on Google's pricing page on 13 November 2026.

Did OpenAI models stop working in Cursor on 12 November?

As of 13 November 2026 there is no public confirmation that they did. Cursor's models page still listed GPT models when checked that day, and Cursor had published no announcement. OpenAI called 12 November a proposed shutoff date, so treat GPT access in Cursor as something that can end at short notice.

Does the lawsuit to freeze AI development affect the models I use today?

Not today. The suit was filed on 12 November 2026 in San Francisco Superior Court and asks for an order barring further development of the most advanced models. No court has ruled on it, and the coverage read for this post mentions no hearing date and nothing about access to models that are already released.