Gemini 4 Hydrogen Pricing, Access, and Claude Meetings Beta

Two product posts landed on November 12, 2026, and they are not the same kind of release. Google’s is a model, with a published price, a context window, and a list of surfaces. Anthropic’s is a meeting feature, with a beta plan limit, a default privacy rule, and a date when in-call artifacts stop being free. Reading them as one “AI news day” hides the part a buyer can actually check.

Carbon, which Google had put in front of Ultra subscribers two days earlier, is the context for Hydrogen. It is not a second full specification in this piece, because the Carbon announcement link on the company post resolved to a video rather than a blog, and the facts below about Carbon are limited to what the November 10 posts say plus what the Hydrogen blog itself repeats.

What shipped, and when the clocks say it shipped

Demis Hassabis posted at 18:11 UTC on November 12 that Gemini 4 Hydrogen was out, and pointed at the Google blog. Logan Kilpatrick’s quoted post, six minutes earlier, said the model was available to all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity. The Gemini app account, at 18:06 UTC, said the rollout was global on desktop, mobile, and web, and that free users would get Hydrogen through auto model selection when it was the appropriate model for the prompt.

The blog page, by Koray Kavukcuoglu, lists a last-updated date of November 13, 2026. The body says Hydrogen is being introduced “today” and is rolling out “starting today.” The X posts are November 12 UTC. The page date is November 13. This article uses the page date for the written specification and the post timestamps for the public announcement. It does not collapse them into a single hour.

Claude’s clock is cleaner. The @claudeai post at 20:16 UTC on November 12 introduced Claude Meetings as a beta on Team and Enterprise plans. The follow-up, two seconds later, said it starts with Google Meet, with other platforms coming soon, and linked the product article. That article is dated November 12, 2026.

Carbon’s posts are November 10. Sundar Pichai wrote that Gemini 4 Carbon was rolling out that day for sustained, long-horizon tasks, reasoning through complex software engineering, cybersecurity, and scientific discovery, coming globally via the Gemini API in Google AI Studio and Gemini for business, and for Ultra. The Google account called Carbon the strongest model, for long-horizon tasks and multi-step reasoning, and said that if Argon is for complex tasks, Carbon is the model for the hardest problems. The Gemini app account said the frontier model was rolling out to Google AI Ultra subscribers on desktop, mobile, and web at gemini.google.

Access is already a split. Carbon, on those posts, is an Ultra rollout. Hydrogen, on the November 12 posts and the blog, is the broad one, with a footnote that free Gemini app users see it through auto model selection rather than as an always-on picker choice.

The price and the window are the part you can audit

The Hydrogen blog prices the model flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens. “Flat” is the operative word. The page does not publish a cheaper short-context tier or a surcharge past some smaller window. A 1 million-token input is the same input rate as a short prompt, on this page.

Output is a separate meter, and the page also says Hydrogen pairs native 1 million-token input with up to 1 million output tokens for multimodal knowledge work: dense PDFs, spreadsheets, charts, audio, and recorded video, turned into reports. A million output tokens at $7.50 per million is $7.50 if a job actually emits that much. Most jobs will not. The rate is still the number to put in a calculator, not a blended “about a dollar” paraphrase.

Thinking level is the other cost lever the page names, and it does not attach a separate price to each level. Hydrogen has four: low, medium, high (the default), and a new xhigh tier. The blog says they scale quality, latency, and cost so you wait and pay for the reasoning depth a task needs. It does not print a multiplier for xhigh versus high. Anyone budgeting xhigh from the $1.50 / $7.50 card alone is guessing. The card is the flat token rate. The thinking-level cost difference is described, not tabulated.

On the Artificial Analysis Intelligence Index, the blog says Hydrogen surpasses Argon at the default high thinking level, 51 versus 50, and reaches 54 at the highest xhigh setting, while often finishing faster than models in its class. A one-point gap at the default setting is what the page prints. It is not a claim that Hydrogen replaces Argon on the long-horizon work Pichai assigned to Carbon two days earlier. The blog keeps that split explicit: Carbon and Argon excel at deep, long-horizon reasoning on the hardest problems. Hydrogen is built for rapid execution across agentic workflows.

Where Hydrogen sits relative to Carbon, in Google’s own words

The practical architecture Google describes is multi-model, not a single winner. In real software work, the blog says teams still use frontier models such as Carbon and Argon for architectural judgment and subtle bugs, and still need a fast execution model for writing code, running tests, and checking changes. Hydrogen is meant to stand alone for everyday work, or to pair with Carbon and Argon as the execution engine.

Orchestrator, in Google Antigravity, is the mechanism they name for making that switch automatic. Rather than one model from start to finish, Orchestrator moves between models such as Carbon and Hydrogen in the same session. On Terminal-Bench 4.0, the blog says Orchestrator reached near-Carbon task success while, on average, using 24 percent less quota and finishing 19 percent faster. For parallel agents, Carbon can coordinate while Hydrogen runs worker and verifier agents in isolated workspaces.

Those percentages are averages on one named bench, for the orchestrated setup, not a promise that every repository will be 19 percent faster if you add Hydrogen to a Carbon session by hand. The page does not say Orchestrator is on by default for every Antigravity user. It says they are introducing it.

The standalone comparison is against Gemini 3.8 Flash, not against Carbon. At the high default setting, Hydrogen finishes tasks 2.5 times faster and more than triples Terminal-Bench 4.1 scores, 57 percent versus 18.2 percent. It cuts fabricated claims by more than 5 times, and the blog says it adapts when a tool call fails or a search returns empty. Partner early testing on complex graphs, SQL, and data-heavy enterprise workflows saw Hydrogen run 2 times to 2.8 times faster than 3.8 Flash, with token use down 12 percent to 23 percent. “Partners have seen” is the attribution. It is not a Google-wide median printed as a guarantee.

Other benches on the page, each with the scope the sentence gives them: number 5 on Agent Arena, arena.ai’s live benchmark, for multi-step tasks. Leads its class on HLE-Diamond at 58.8 percent. Number 1 overall on Vals Tax Agent Bench at 52.3 percent. Tied for number 1 on the Vals Legal Research Bench at 55.8 percent. Leads its class on AutomationBench-AA. Multimodal Knowledge Agent benchmark at 53.4 percent. Agents’ Last Exam v1.1 at 71.6 percent. SWE-sweep is cited as a win over similar-tier models for finding and fixing bugs without breaking existing code, without a printed score. The page also says the model batches file-read and search calls instead of opening files one by one. That is a behavior claim, not a benchmark row.

A buyer who only remembers “54 on the index” has skipped the migration story. The 2.5 times speed and the Terminal-Bench 4.1 jump are the sentences aimed at teams leaving 3.8 Flash. The one-point edge over Argon is a different comparison, at a named thinking level. Carbon remains, in the November 10 posts, the strongest model and the Ultra-gated one.

Who can call it on Monday

The blog’s availability list matches Logan Kilpatrick’s post and adds a few product edges.

Developers build with the Gemini API in AI Studio and Android Studio, and paid Google AI Pro or Ultra subscribers can try Hydrogen in Google Antigravity. Enterprises deploy it in Gemini for business through the API or the model dropdown, in Antigravity, and in Gemini in Google Sheets for eligible business subscriptions. Everyone can start in the Gemini app model dropdown and in Google Search. Gemini in Sheets starts with Pro and Ultra. The footnote repeats the free-tier rule: in the Gemini app, free users get Hydrogen via auto model selection when it is the appropriate model for the prompt.

Search gets its own section. The blog says Hydrogen makes Search more able to take a question in a natural run of thought, including multi-part queries, hard-to-read typos, and specific requests. Responses are described as more visual. Query fan-out is being upgraded so Search connects to a broader set of links. None of that is a ranking-factor change a site owner can implement from this post. It is a model swap inside Search, dated to the same rollout.

In the Gemini app, the blog says Hydrogen follows guidance more closely, calls tools more often, and needs less back-and-forth. It can pull context from uploads, from Gmail and Drive, and from connected apps including Airtable and Dropbox. In Sheets, it says outputs come back over 30 percent faster than previous models in the Sheets canvas. That 30 percent is a Sheets-canvas comparison, not the 2.5 times developer-task figure, and not the 19 percent Orchestrator figure. Three speed claims, three baselines. Quoting one of them as “Hydrogen is 30 percent faster” drops the baseline and misstates the page.

Safety numbers are on the same post, and they are internal or named-bench figures, not a third-party audit attached as a PDF in the article body. The blog says Hydrogen reduces overreach in agentic tasks by up to 81 percent and under-service by up to 79 percent, and cuts misaligned actions and task shortcutting across tool-use and coding evaluations by 12 times, relative to the leap it describes over 3.8 Flash’s safeguards. On impossible or blocked tasks it is up to 2 times less likely to bypass permission boundaries, with a 94 percent honesty rate on an internal agent honesty evaluation. Chain-of-thought legibility is preserved at 99.9 percent, with zero deceptive reasoning edits across more than 200,000 audited turns. On the Gray Swan indirect prompt-injection benchmark it is more than 2.3 times more resilient. False refusals on HealthBench Pro are under 0.4 percent. The model card is referenced for more detail. This article does not invent card tables that were not in the blog text.

The consumer examples on the page (a voxel island, a storefront shader, a spreadsheet pixel canvas, a slides audit, a lightning globe) are demonstrations. They are not benchmarks. Leaving them out of a buying note is not a gap in the spec. The spec is the price, the window, the thinking levels, the surfaces, and the benches with numbers.

Claude Meetings is a participant, not a model card

Anthropic did not publish a token price on November 12. It published a behavior and a set of defaults. Claude can join a video call as a visible participant, follow the conversation, and start work during the meeting. The beta is on Team and Enterprise plans. Support is currently Google Meet. Other platforms are “coming soon,” which is the article’s phrase, not a date.

People on the call can ask questions out loud or in chat. By default Claude answers in writing. It can use connected tools with permission. It can build docs, whiteboards, or other artifacts while teammates on the invite edit them. It writes the notes and the action items. After the call, notes, action items, transcript, and anything it built sit on that meeting’s page in claude.ai. You can later ask Claude about meetings you were part of, or add it to the next one from the calendar in claude.ai.

The article’s customer-event scene is an illustration of that flow, not a measured deployment. There is no latency number, no accuracy rate for notes, and no count of teams in the beta. Treating the illustration as a case study would add facts the page does not have.

What the page does specify is the privacy default, and that default is the part worth reading twice. Claude is set to stay out of calls when someone from outside the organization is invited, and to leave if an outsider joins partway through. An admin can allow outside guests. On those calls Claude posts a notice in chat and does not reply there, while what guests say is still included in the notes and transcript. That is a narrower silence than “the meeting is private.” Guests are in the record. Claude is just not chatting with them.

Tool use is per person. Claude uses each person’s connected tools with that person’s permission, asks the first time, and treats a yes as applying going forward. If the request is spoken, it asks each time to confirm who is speaking. An answer posted in the call chat is visible to everyone on the call. When it joins, it appears as a participant and posts a notice. It saves a transcript, a summary, and screen descriptions. It does not save audio. The article tells admins to make sure everyone on the call is willing to be transcribed before auto-join is turned on.

Plan controls differ. On Team plans, inviting Claude is available by default and an admin can turn it off. On Enterprise plans, an admin turns it on for all or part of the organization. During beta it is not available to some organizations, including those with zero data retention, a HIPAA-ready configuration, or customer-managed encryption keys. That exclusion is a buying constraint, not a footnote. A hospital deployment on a HIPAA-ready setup, or a company that required CMEK, cannot treat the November 12 post as “Meetings is on.”

The price of the beta is also specific, and it expires in one place. Notes and transcripts are free on Team and Enterprise during the beta. Answers in the call are free unless they use a connected tool, in which case they count toward plan usage. Docs, whiteboards, and other artifacts made in a call are free through February 10, 2027. After that date the page does not say what artifacts cost. “Free through February 10” is not “free.”

Partner names are a roadmap, not a connector list you can click today. Anthropic says it is working with Gong, Attention, Granola, Fireflies.ai, Otter, and Plaud so a meeting transcript can soon be sent to tools a team already uses. “Soon” is doing the work in that sentence. The live path described now is the meeting page inside claude.ai, plus Google Meet.

Two rollouts, two questions

Hydrogen asks a procurement question. At $1.50 and $7.50 per million tokens, flat across a million-token window, with four thinking levels and a free-tier path that is auto-selection rather than a pinned model, the open item is which surface a team is actually on. API and AI Studio are in the “all tiers” sentence. Antigravity for consumers is paid Pro or Ultra. Sheets for the Gemini side starts at Pro and Ultra, with a separate “eligible business subscriptions” line for enterprise Sheets. Search and the app dropdown are the broad consumer surfaces. Carbon, two days earlier, was the Ultra frontier post. Using Hydrogen as if it were that frontier model contradicts both the November 10 wording and the Hydrogen blog’s own split.

Claude Meetings asks an admin question. Is the workspace Team or Enterprise, is Meet the conferencing tool, is an outsider ever on the invite, and does the org run ZDR, HIPAA-ready settings, or CMEK? If the answer to the last of those is yes, the beta exclusion applies. If outsiders join, the default is that Claude leaves, unless an admin has allowed guest calls, in which case guests are transcribed and Claude does not chat in the guest-visible thread. Artifact fees start being a question after February 10, 2027. Until then the meter that already runs is connected-tool usage on spoken or typed answers.

Neither post is a reason to treat November 12 as a single platform shift. One is a priced model with a published index gap of one point over Argon at the default thinking level, a much larger jump versus 3.8 Flash on Terminal-Bench 4.1, and a routing feature that tries to spend Carbon only where the blog says Carbon still matters. The other is a visible meeting participant with a written-answer default, no saved audio, and a guest rule that is easy to miss if you only read the launch sentence about docs and whiteboards.