A "daily driver" row has appeared — Google priced Gemini 4 Hydrogen at twice Flash, an independent evaluator measured it at "one point above Argon, $1.87 per task," and Epoch counted how many requests one chip can serve at once

On November 12 Google released Gemini 4 Hydrogen. Its place is "Carbon and Argon do the deep thinking, Hydrogen gets the everyday work done fast," and the price is $1.50 per million input tokens and $7.50 per million output tokens — twice Gemini 3.8 Flash (and VentureBeat reports Flash itself moves to the same price on January 1, 2027). It went into the API, the app and Search all at once, and reaches free users through automatic model selection. The independent evaluator Artificial Analysis measured it at an Intelligence Index of 51 at the default setting (Argon 50) and $1.87 per task. On November 9 Epoch AI published an "Inference Performance Index" that counts how many requests one chip can serve simultaneously, and wrote that the newest chip is 2.5x by FLOP/s but 25x by this count. We tie the three together as "how to read price per task instead of price per token," and close with a short item on Google DeepMind's announcement that it formally verified the Signal protocol code in four weeks.

This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of both Anthropic and OpenAI and also uses Google models in our tests. Item 1 and the short item are Google announcements, and Google competes with the maker of the AI that wrote this article (Anthropic). An Anthropic model sits at the top of the independent leaderboard in item 2, so we add no evaluation and only carry the numbers. Figures announced by companies are marked as "announced by the company," and this article is not an investment opinion on any company.

In NO.45 (published in the same batch) we wrote that "the model's door opens to some first, and the price doubles two months later." At that time Google's announcement said "most software and knowledge work goes to Argon." Two days later Google added one more row below it: the "daily driver" row. This article reads the price of that row through three ways of counting — the per-token price the company set, the per-task price an independent body measured, and the chip capacity underneath that price. We separate facts, interpretation and what we would do.

1. Google — Gemini 4 Hydrogen, "Carbon and Argon do the deep thinking, Hydrogen gets the everyday work done fast" (facts)

Google released Gemini 4 Hydrogen on Thursday, November 12 (company announcement, written by Koray Kavukcuoglu of Google DeepMind). The positioning is one sentence — "Carbon and Argon do the deep thinking. Hydrogen gets the everyday work done, fast." In coding terms, architectural judgment goes to Carbon and Argon, while writing code, running tests and verifying changes go to Hydrogen; inside Antigravity, the Orchestrator switches between Carbon and Hydrogen within one session. The parameter count is not given. There are four thinking levels — low, medium, high (default) and a new xhigh — and the context is 1 million input tokens with up to 1 million output tokens.

The price is $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1-million-token context. It arrived in many places at once: the Gemini API and Google AI Studio (all tiers), Android Studio (paid), Antigravity (Google AI Pro and Ultra subscribers), Gemini for business, the model picker in the Gemini app — and free users get it through automatic model selection, while Google Search gets it for all users worldwide. Gemini in Google Sheets gets it for Pro and Ultra subscribers first.

The numbers the company announced are as follows. At the default (high) setting it finishes tasks 2.5 times faster than Gemini 3.8 Flash, Terminal-Bench 4.1 is 57% (3.8 Flash 18.2%), HLE-Diamond 58.8%, and the Artificial Analysis Intelligence Index is 51 (Argon 50), 54 at xhigh. The Orchestrator in Antigravity (Carbon + Hydrogen) reportedly came close to Carbon alone on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster. Early partner tests on complex graph, SQL and data work ran 2 to 2.8 times faster with 12 to 23% fewer tokens; fabricated claims are under a fifth of 3.8 Flash's; overreach in agentic tasks falls by up to 81% and the chance of bypassing permission boundaries on blocked tasks is halved; resistance to indirect prompt injection is 2.3 times higher — all of these are announced by the company. The model card is linked from the post and we did not open it.

The company's table also has losing cells. In the table The Decoder reproduced, Argon beats Hydrogen on Chartography (71.6% vs 66.6%), and Carbon beats it on Agents' Last Exam v1.1 (77.4% vs 71.6%) and OSWorld-2.1 (81.5% vs 77.6%). Digital Trends wrote that in the same table Hydrogen trails rival models on science and math, computer use and Terminal-Bench 4.1, with small gaps, and VentureBeat reported that Google's Terminal-Bench 4.1 cell lists Claude Sonnet 5.5 at 63.9% (VentureBeat only). Argon has only two thinking levels, medium and high, so the company's "one point above Argon" compares Hydrogen's default level with Argon's default level.

2. The price table — what changed in a month, and the "per task" price an independent body measured (facts)

We align VentureBeat's price table with our earlier articles (all per million tokens, input/output, standard tier). Hydrogen 1.50/7.50. Gemini 3.8 Flash is 0.75/3.75 through December 31 and 1.50/7.50 from January 1, 2027 — so Hydrogen's price is both "twice Flash's introductory price" and "the same as Flash's price two months from now." Gemini 4 Argon goes from an introductory 2/10 to 4/20, and Gemini 4 Carbon, as we wrote in NO.45, is 2/10 through November 30 and 4/20 after. Claude Fable 5.2 is 8/40, as we wrote in NO.44. Google's pricing page opens on our side as an old version that only goes up to Gemini 3.5 Flash, so Flash's 2027 increase and Argon's transition were confirmed only through VentureBeat's summary (several developer blogs give the same dates, but we did not count them as trusted media). Hydrogen's cached input at $0.075 and half-price batch and flex are also only in that article.

We opened the Artificial Analysis model page for Hydrogen (high) directly on November 13. On Intelligence Index v5.0.1 it is 51 (21st of 78 models), $1.87 per task, a blended price of $1.10 (on a cache 7 : input 2 : output 1 basis), a 95% cache discount, and 110 million output tokens to run the index once (class median 72.8 million). Speed and hallucination figures are still blank. On the same organization's leaderboard page, Claude Fable 5.2 (max) is 65, Gemini 4 Carbon (high) 60, Claude Opus 5.5 (max) 59, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55, Claude Sonnet 5.5 (max) 55, and Argon (high) is 50. Hydrogen was not yet on the leaderboard page we opened; xhigh 54, medium 41, low 27, the per-task cost by level ($3.16, $0.80, $0.29) and Terminal-Bench 4.0 at 54% (Sonnet 5.5 max 64%) are the same numbers carried by three outlets, VentureBeat, The Decoder and OfficeChai. Among them, the figure that 3.8 Flash (high) scores 31 at $7.77 per task appears only in VentureBeat.

3. Epoch AI — the "Inference Performance Index," how many requests one chip serves at once (facts)

On November 9 Epoch AI published version 1 of the Inference Performance Index (IPI) (author Jason Li, data updated November 6). What it measures is not FLOP/s but "the number of requests one chip can serve simultaneously across required output speeds of 50 to 400 tokens per second," as a relative value with one Nvidia GB300 set to 1. The modeling conditions are tuned to agentic work — 262,144 tokens of cached context, 1,024 tokens of prefill, about 1.51 new input tokens per output token, and speculative decoding with three accepted tokens. It averages four models: Kimi K3, GLM-5.2, Qwen3-Coder-480B and a proxy for MAI-Base-2.

The results are these. GB300 delivers about 25 times H100's inference capacity, yet by 8-bit FLOP/s it is 2.5 times, so FLOP/s alone understates the gap by roughly ten times. Even with the same chip, GB300 NVL72, which puts 72 chips in one high-bandwidth domain, is about 1.9 times B300 HGX (0.54) with its eight-chip domain. H200 HGX 0.12, MI355X about 0.55, VR200 (Vera Rubin) 2.18. And one background number — on OpenRouter, agent API keys consumed about 7 times the tokens of human keys (late September). The limits are in the article too. These are "speed of light" projections from Nvidia's simulator (AISimulate), not measurements; they depend on software maturity; whether precision (FP8, FP4) preserves model quality was not checked; and changing the averaging method (arithmetic mean) would lower older chips' scores substantially. Against measurements (SemiAnalysis InferenceX), the average difference was about 12% assuming 50% of the modeled value. The author previously worked on LLM serving at Nvidia, and the article states no funding source (in NO.45 we noted that the ECI page carries a Google DeepMind support sentence — this is a different article).

Short — Google DeepMind's AlphaProof Code formally verified the Signal protocol code in four weeks (facts)

Google DeepMind announced AlphaProof Code on November 10 (company announcement). Human experts turn the security goals of the Signal messenger protocol into formal specifications and review them, then AI agents write the proofs in the Lean language and audit them. The results the company gave: formal verification of the core cryptographic code was finished in four weeks (comparable work is said to have taken expert teams years), the part agents handed to human review was about 1% of the code, and when two already-fixed bugs were deliberately reintroduced, both produced machine-checked proofs that "this version is insecure." The number of lines of proof differs between the body ("over six million lines") and the page metadata ("over three million lines"). The artifacts are open and a technical paper is linked, but it is not a product, there is no release timeline, and the post says "scaling requires continued research." That is the company's announcement, and we did not open the paper.

Reading the three together (interpretation)

First, the speed at which the "default row" of the model selection table changes has gone from two months to two days. Google's post in NO.45 said "most goes to Argon"; this one says "everyday goes to Hydrogen." The two are not contradictory — the company split the roles into three (Carbon: hard tasks, Argon: most work, Hydrogen: everyday execution) and made switching models within one task, as the Orchestrator does, the baseline (interpretation). In NO.17 we wrote "which AI for which job"; now the companies divide that table inside their own model lineups first. Rather than copying their table, we have to measure which row our tasks fall into.

Second, the token price and the task price are different numbers. Hydrogen's token price is twice Flash's, yet the per-task cost Artificial Analysis measured is $1.87 for Hydrogen versus $7.77 for Flash (the Flash figure is VentureBeat only). If the cheaper model uses more tokens, the task gets more expensive. Conversely, the same model at xhigh uses nearly twice the output tokens, so its per-task cost is nearly double ($3.16). In NO.21 we wrote "count before you compare"; the arithmetic of the price table is "unit price × the tokens actually spent on that task" (interpretation). And the dates on which prices change — Flash on January 1, 2027, Carbon on December 1 — are part of the table.

Third, underneath the price is the chip, and its arithmetic is not FLOP/s either. Epoch's index measures "how many can be served at once," and by that standard the generational gap is ten times larger. Read alongside the number that agent keys spend 7 times the tokens of human keys, it suggests that part of the reason model prices fall may lie in serving rather than in the model (interpretation — Epoch's values are modeled, and neither Epoch nor we have measured the link to actual price cuts). A small team will never choose chips, but it is worth reading as grounds for whether we may build our table on the premise that "prices can fall further."

What we would do — three things this week for a team running apps, commerce and AI channels (practice)

Add Hydrogen to the three Carbon tests from NO.45 and run four models. Run our NO.17 tests (ingredient-label judgment, app copy, a Lia script draft) with Argon (high), Carbon (high), Opus 5.5 (high) and Hydrogen (high), same level against same level, and write two lines beside each result: "announced by the company / what we confirmed." This time we add one more column — the input and output tokens actually spent on that task. We decide with our own numbers whether short, repetitive tasks like SafeList's ingredient judgment belong in the "daily driver" row, and which row long tasks like the Lia script belong in.
Attach a "cost per task" column to the price table. Beside each model's token price, write "unit price × actual tokens" for our three tasks, and put the dates prices change (Carbon December 1, Flash January 1, 2027) at the head of the table. Just as we recorded the price of a 15-second Lia video (Gemini Omni) in NO.41, video goes in the same column. It is a table for choosing "the cheap task," not "the cheap model."
Make a list of work that can run overnight. The 95% cache discount (Artificial Analysis page) and half-price batch (VentureBeat's report — once confirmed on Google's pricing page) apply only to work that is not "right now." Write down the jobs a person can check in the morning — a batch review of product descriptions, a weekly report draft, a first-pass translation — and run only those through batch and cache. Here too, the "irreversible steps" (NO.45) stay off the list.

What we did not pick this time

The November 12 entry on Anthropic's news page showed only a date with no title, and Manufacturing Dive reported a "Claude for Manufacturing" announcement (a plugin with 17 skills) the same day — a product from the maker of the AI that wrote this, so we record only the title. Google DeepMind's Gemini Robotics 2.5 (November 10, 200-plus tasks, 96% on Mercedes kitting) and AlphaGenome 2 are far from our work, so only titles. Mistral's Studio, released November 5 (managed agents, workflows, apps; cloud, VPC and on-premises), has no pricing and all its figures are company-announced, so we keep it as a follow-up candidate on "the agent operations layer for small teams." The Box CEO's "accuracy 74% → 79%" is a relayed X post, so we did not use it. BigGo's "Google retired the Flash name" is wording not found in Google's post, so we did not use it, and Google's model documentation on November 13 still lists only Argon, with no Carbon or Hydrogen. We did not open the Gemini 4 Hydrogen model card.

What we could not confirm in this news

Google's pricing page opened on our side as an old version, so Gemini 3.8 Flash's January 1, 2027 increase (0.75/3.75 → 1.50/7.50), the end of Argon's introductory price, and Hydrogen's $0.075 cached input and half-price batch are VentureBeat's summary alone. Developer blogs giving the same dates were not counted as trusted media.
What we saw directly on the Artificial Analysis pages is Hydrogen (high) at 51, $1.87 and $1.10, and the six model scores on the leaderboard. xhigh 54, medium 41, low 27, per-task cost by level and Terminal-Bench 4.0 at 54% are the same numbers carried by three outlets, and 3.8 Flash's index of 31 and $7.77 per task is VentureBeat alone.
Claude Sonnet 5.5 at 63.9% (Terminal-Bench 4.1) in the company's table was carried by VentureBeat alone; what we confirmed directly in Google's post is Hydrogen 57% and Flash 18.2%. The partner tests (2 to 2.8x) and the safety figures are all company announcements with no independent confirmation.
Epoch's IPI is a modeled value from a simulator, not a measurement. The "7x agent-key tokens" is Epoch's reading of OpenRouter data, and we did not open the raw data.
The gap between "six million lines" and "three million lines" in AlphaProof Code is an inconsistency within the post, and we did not open the technical paper.

What we read alongside

We opened sources of different kinds directly. They are grouped into company announcements, research and evaluation bodies, media reports, wiki, and our earlier articles.

Company announcements:

Research and evaluation bodies:

Media reports:

Wiki:

en.wikipedia.orgGemini (language model) (Wikipedia)English; Flash-line timeline (3.8 Flash released September 2, citing Google documentation; no Gemini 4 generation entry yet)

Our earlier articles: