How much does Gemini 4 Hydrogen cost?

Google's blog and pricing page list $1.50 per million input tokens and $7.50 per million output tokens, and the blog says the price is flat across the model's 1 million-token context window. The pricing page says the output price includes thinking tokens. Google's Gemini API release notes, dated November 12, 2026, label the model generally available (GA), and Google Cloud's page also prints "Launch stage: GA"; the blog prints no stage label. These are list prices from one table, not a forecast of any bill, and we repeat none of Google's benchmark claims.

What did AWS change in the AgentCore runtime?

AWS's What's New post of November 12, 2026 says new runtimes in Amazon Bedrock AgentCore now use the next generation of the runtime by default in 19 listed Regions, while existing runtimes stay on the current version until moved. AWS says sessions start in under 2 seconds at P75 on a 2 GB image, down from 29.7 seconds, and that multi-turn agents can see up to 68% less billed memory per session. AWS's pricing page lists higher per-unit rates for the new runtime than for the original (the table in the article has the numbers), and AWS's November 9 roundup says most agents should see a lower bill, though that is not guaranteed.

What did Holiday Extras report after rebuilding its app with Codex?

OpenAI's customer story, dated November 12, 2026, says open-to-book conversion on the main app increased 32.5% following the rebuild and the error rate fell from 19% to 9%. It says the team had a fully working rebuild within two weeks, progress it says would previously have taken over six months. The page's own caveat is that results compare the previous and rebuilt app across different periods, and the page says the rebuild also rethought the experience, that OpenAI models power the app's personalised experience and that ChatGPT Enterprise was part of the rollout. It is a story published by OpenAI, which sells Codex.

Is Databricks Genie ZeroOps generally available?

The pages we read label it Beta, not generally available. Databricks' release notes, dated November 12, 2026, say it is available in Beta, and its docs say a workspace admin must turn it on from the Previews page. Status as of November 12: Beta, and releases are staged. The docs say it does not change production data or code without human review. The pages we read state no price for ZeroOps itself; the overview adds that detecting data quality issues needs separately billed data quality monitoring.

Do shoppers let AI agents check out for them?

Salesforce's survey story, dated November 12, 2026, says 19% of personal agent users have deployed their agents at checkout. That is a share of a subgroup whose size the story does not print; the story cites a survey of more than 2,500 shoppers in four countries. The story and the two Salesforce pages it links to do not state who ran the survey, when, or the margin of error. Salesforce sells commerce and agent products.

Signal / Noise Nov 12: Gemini 4 Hydrogen, AgentCore

An upright glass panel in a brushed-metal frame stands on a dark glossy floor. Across the top is a row of five blank glass tiles, the fourth lit solid red. Along the lower half, a tangle of dim grey jagged lines on the left smooths into one clean glowing red wave that rises under the red tile.

The short version

Five AI notices from November 12; for each we say whose numbers they are. Google prints $1.50 input and $7.50 output per million tokens for Gemini 4 Hydrogen, and its API release notes label it generally available. AWS made its next-generation agent runtime the default for new runtimes; its pricing page lists higher per-unit rates. Also: a vendor-published customer story and shopper survey, and Databricks' Genie ZeroOps (Beta).

In plain words

Think of this edition as five notices pinned to one board. One is a price list for a new Google model, and one is an AWS update to the place where agents run. One is a customer story published by the company that sells the tool, and one is a survey published by a company that sells commerce and agent software. The last is a Beta tool that watches data jobs and suggests fixes. For each we say in one plain sentence what it is, whose numbers these are, and what we could not check.

What does an AI agent cost to run, and how far should you trust a vendor's own numbers? Each of today's five items bears on one of those two questions.

This edition has one Signal item and four Watch items. Signal means you can act on it this week. Watch means worth knowing; one quick check at most. Most sources are dated November 12, 2026; the Sources list marks the older and the undated ones. Every number belongs to whoever published it. The "What it means for you" and "Our take" lines are ours. For the previous edition, see the November 11 edition.

Signal: What does Google list as the price of Gemini 4 Hydrogen?

Google says Gemini 4 Hydrogen is built for fast, everyday coding and agent work. Its blog post, dated November 12, 2026, says the model is "priced flat across its entire 1 million-token context window" at $1.50 per million input tokens and $7.50 per million output tokens. A token is a chunk of text, often part of a word. A context window is how much text a model can take in at once.

What stage does Google call it?

Google's Gemini API release notes, in an entry dated November 12, 2026, label gemini-4-hydrogen "generally available (GA)". The blog post prints no stage label; it says "Starting today, Hydrogen is rolling out globally" across its developer, enterprise and consumer surfaces. Google Cloud's page for the model on its Gemini Enterprise Agent Platform prints "Launch stage: GA". We read GA as the label for the model in Google's developer API and on Google Cloud's platform, not for the Gemini app or Search.

Status as of November 12, 2026: GA in the API release notes and on the Google Cloud page; "rolling out globally" in the blog.

How does Hydrogen's price compare with Gemini 3.8 Flash?

Google's blog calls Hydrogen "a major standalone upgrade for teams migrating from Gemini 3.8 Flash". Here are both models' rows from Google's pricing page (Standard tab, paid tier, per million tokens):

	Input	Output (includes thinking tokens)
Gemini 4 Hydrogen	$1.50	$7.50
Gemini 3.8 Flash, through Dec 31, 2026	$0.75	$3.75
Gemini 3.8 Flash, starting Jan 1, 2027	$1.50	$7.50

Our arithmetic: $1.50 / $0.75 = 2.0 and $7.50 / $3.75 = 2.0, so Hydrogen's printed prices are double the 3.8 Flash rows through December 31, 2026, and equal to them from January 1, 2027. It is not a forecast of any bill.

The blog, the release notes and the pricing page print no end date for Hydrogen's price. Our November 11 edition covered Gemini 4 Carbon's price step; we do not compare the two models.

What it means for you: if a product you pay for runs on a Google model, ask which model and which price row applies to you. Per-million-token prices are only one input to an agent's monthly bill; our teardown of Google Cloud's cost estimator agent lists the others.

Our take: Signal, because the price is printed and the Flash rows change on January 1, 2027, so a vendor's answer can be checked against a number this week.

Watch: What changed in AWS's default AgentCore runtime, and what does it cost?

AWS says its updated agent-hosting service starts sessions faster, lets them be larger, and shrinks the memory it bills while an agent waits. AgentCore is AWS's service for running AI agents; its runtime hosts each agent session. AWS's What's New post of November 12, 2026 says that starting that day, new runtimes "use the next generation of AgentCore runtime by default". Existing runtimes "stay on the current version until you move them".

What is new, and what is not?

The next generation itself is older: AWS's post of September 18, 2026 announced it, including "scale to zero". What is new on November 12 is the default, in 19 listed Regions. AWS's claims:

Sessions up to 8 vCPU (virtual processor cores) and 32 GB of memory.
On a 2 GB image, sessions start in under 2 seconds at P75 (the time within which three in four starts finish), down from 29.7 seconds on the previous generation.
For multi-turn agents, "up to 68% less billed memory per session". That figure covers memory only, for agents "with natural pauses between interactions"; it is "up to", not a bill.

What do AWS's own pages say about the bill?

The November 12 blog post says the runtime uses compute more efficiently, so "you pay less for the same work", and neither November 12 post prints a dollar rate. AWS's pricing page does, and it lists higher per-unit rates for the new runtime:

AgentCore pricing page, runtime	CPU, per vCPU-hour	Memory, per GB-hour
Original runtime (v1)	$0.0895	$0.00945
Next generation (v2), consumption	$0.1276	$0.0169
Next generation (v2), committed baseline (availability unclear, see below)	$0.0997	$0.0132
Our arithmetic: v2 consumption against v1	42.6% higher (0.1276 / 0.0895)	78.8% higher (0.0169 / 0.00945)

AWS's weekly roundup of November 9, 2026 says the per-unit rate is higher and the team expects most agents to see a lower bill, "though that isn't guaranteed".

Do you pay while an agent waits?

AWS's pricing page says CPU scales to zero while an agent waits on models or tools, and idle memory is reclaimed after 120 seconds. It also says a billed session includes idle periods and that billing includes system overhead. It states no separate charge for an idle agent.

What else do AWS's pages leave unsettled?

The pages disagree on whether committed baseline pricing is available: the blog describes it in the present tense, the pricing page marks those rates "launching by October 2026", and the roundup says "coming soon". We do not say it can be bought. AWS's pricing page has one worked example for the new runtime's rates in the table above (a customer support agent, 1 million sessions a month, under AWS's own assumptions) and none for the original, so that page does not compare the two bills. The November 12 blog quotes customers' own comparisons (one reports "reducing measured runtime cost per call by 17%"); we could not check them. The November 12 pages print no stage label; the roundup says the runtime "became generally available".

What it means for you: in our reading, ask whoever runs your agents on AWS which runtime version they use and which rate row applies. Do not assume AWS's "up to" memory saving outweighs the higher rates until your own usage says so. For how AgentCore compares with two other platforms, see our platform comparison.

Our take: Watch. These pages give a price list, claims, customer quotes and one example under AWS's assumptions, not a side-by-side bill we can check. The answer depends on your usage.

Watch: What numbers does OpenAI's Holiday Extras customer story report?

OpenAI's customer story says a travel company rebuilt its main app with Codex, OpenAI's coding agent, and reports before-and-after figures. It is a page published by OpenAI, which sells Codex. It is titled "Holiday Extras is rethinking what it can build with Codex" and dated November 12, 2026. We read it in a browser and did not check the figures with the company.

What is the page's own caveat?

It comes first: "Conversion and error-rate results compare the previous and rebuilt main app across different periods." The page does not state how long either period was.

What does the page report?

Conversion: "open-to-book conversion on the main app increased 32.5% following the rebuild". The page does not define open-to-book, say whether 32.5% is a relative rise or percentage points, or give the starting rate.
Errors: the main app's error rate "fell from 19% to 9%". The page does not define error rate. Its headline tile shows "10%" for the same change; 19 minus 9 is 10 percentage points (our arithmetic).
Speed: the team had "a fully working rebuild" within two weeks, progress that "previously would have taken over six months". The second number is the page's estimate of what the work would have taken, not a measurement, and the page gives no method.
Releases: the main app moved from an annual major release "to weekly releases in the first month after the rebuild".

Can the gain be put down to Codex alone?

The rebuild also changed the product: the page says Holiday Extras "challenged itself to rethink the experience and how it was built". The rebuilt app "offered a more personal experience", "OpenAI models power the app's personalised experience", and the page lists ChatGPT as well as Codex among the products, saying company-wide ChatGPT Enterprise access "established the foundation". Rethinking the experience, personalisation and the coding tool arrive together in these figures. The page does not state Codex cost, team size or any independent audit.

What it means for you: when a vendor shows before-and-after numbers, ask for the periods, the definitions and what else changed. For a comparison of coding agents, see our coding agents comparison.

Our take: Watch. A vendor's story about one customer, with a caveat the page itself prints.

Watch: What does Salesforce's holiday survey say about shoppers and AI agents?

Salesforce's survey story says shoppers use AI agents for research and advice, and that most are not ready to let one buy. It was published by Salesforce on November 12, 2026. Salesforce sells commerce and agent products, so read the numbers as Salesforce's.

What are the numbers, and who are they about?

The story cites a new Salesforce survey of "more than 2,500 shoppers in the U.S., the U.K., Canada, and Australia". It says 19% of "personal agent users" have deployed their agents at checkout. That base is a subgroup: the story says 31% of consumers have tried a personal AI agent, and it does not print how many people the 19% stands for. So this is not 19% of shoppers. The story does not define "deployed their agents at checkout", though it frames shoppers as "not quite ready to let agents check out for them".

What does the survey leave out?

Neither the story nor the two Salesforce pages it links to (an October 9, 2026 article on personal agents and a July 20, 2026 holiday predictions post) print who ran the fieldwork, when, how respondents were chosen, the margin of error or the question wording. The story also cites a Cyber Week projection that it attributes to "Salesforce proprietary research", not to this survey, and a traffic-growth figure with no source of its own; we repeat neither.

Related: our November 11 edition on United's rules for AI assistants; for agents that buy, see our Visa agentic commerce benchmark decode.

What it means for you: when a vendor cites a survey, ask for the dates, the sample and the question wording.

Our take: Watch. A vendor survey with no method on the pages we read.

Watch: What is Databricks Genie ZeroOps, and what stage is it in?

Databricks says Genie ZeroOps watches the jobs, pipelines and tables you choose, works out why one failed or went stale, and proposes a fix for a person to apply. Its release notes, dated November 12, 2026, say it is "now available in Beta"; a workspace admin must turn it on. Status as of November 12: Beta. "Releases are staged."

What does it read?

Databricks' docs say that selecting assets means Databricks processes them with a large language model "on a continuous, autonomous basis"; for a job that runs code from a Git repository (where a team stores its code) this includes the repository contents. People you share your configuration with see "code excerpts, suggested fixes, and sample data rows" in the insights even without access to the underlying assets. The docs also say it "doesn't change production data or code without human review" and "is not HIPAA compliant" (HIPAA is the US health-privacy law; our gloss).

What does Beta mean here?

The Previews page says a Private Preview or Beta product is "not intended for use in production", has no uptime promise (an SLA; our gloss) and no formal support. It also says that during Private Preview or Beta Databricks may "collect and use data relating to this feature, such as inputs and outputs". ZeroOps is labelled Beta, so in our reading that applies. The pages we read state no price for ZeroOps itself, and the overview says detecting data quality issues on tables "requires separately billed data quality monitoring". This is a different feature from the Genie file analysis we covered.

What it means for you: ask what it costs to run (including the separately billed monitoring), what it reads (including code repositories), who it is shared with and what Beta means in your contract.

Our take: Watch. A Beta tool with a stated data-use sentence and no price on the pages we read.

Do this week

Five quick checks for an owner or engineering lead; this is not legal or buying advice.

If a product you pay for runs on a Google model, ask which model and which price row applies to you.
Ask whoever runs your agents on AWS which runtime version they use and which rate row applies.
When a vendor shows before-and-after numbers, ask for the periods, the definitions and what else changed.
When a vendor cites a survey, ask for the dates, the sample and the question wording.
If a data team is offered a Beta agent, ask what it costs, what it reads, who it is shared with and what Beta means in your contract.

Book an AI discovery call

How we know, and what we could not check

We read every source on November 12, 2026 and tested nothing. The OpenAI story and Google's release notes and pricing pages were read in a browser; the other pages as page text. The Gemini and AWS pricing pages print no date. Not read: the other Gemini price tabs, Google's page on what is new in Hydrogen, AWS's AgentCore documentation, Holiday Extras' own site, Salesforce's raw survey data and Databricks' Genie pricing table. For help placing agents in your business, see our AI agents work.

Signal / Noise (Nov 12, 2026): Gemini 4 Hydrogen's Price, AgentCore's New Default, a Codex Rebuild Story, a Shopper Survey, Genie ZeroOps