Google says Gemini 4 Hydrogen beats Argon 51 to 50 on one index; its model card prints no cyber figure

Hydrogen leads Argon by one index point (derived) at 38 percent of Argon's cost per task (derived)

On 12 November 2026 Google published Introducing Gemini 4 Hydrogen (its page data gives published 18:00 UTC and last modified 21:03 UTC the same day). It says Hydrogen "surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level (51 vs. 50, and reaching 54 at its highest xhigh thinking setting)". The gap is 1 point at high and 4 points at xhigh (derived: 51 minus 50; 54 minus 50). This claim can be checked against an independent page: Artificial Analysis's own pages list Hydrogen at High as 51, Hydrogen at xhigh as 54 and Argon at High as 50. The story brief for this briefing listed independent reproduction of 51 and 54 as not established; the saved pages show it is established, on that one index.

What the independent page adds is the cost the launch post leaves out. Artificial Analysis prints $1.87 per Intelligence Index task for Hydrogen at High, $3.16 at xhigh, and $4.97 for Argon at High. Hydrogen at High costs 38 percent of Argon per task (derived: 1.87 divided by 4.97 = 0.376), and Artificial Analysis times it at 7.9 minutes a task against 30.0 for Argon, 3.8 times faster (derived: 30.0 divided by 7.9).

What this does not establish. It does not show that Hydrogen can do what Argon can do on your tasks. The index is one composite of nine evaluations. On the Vals Index, where Hydrogen was run at xhigh, it sits 0.90 points below Argon (derived: 68.49 minus 69.39), a gap inside the 0.99 that Vals prints beside Hydrogen's score. Carbon is 9 index points above Hydrogen at High on Artificial Analysis (derived: 60 minus 51). One point is also not a ranking: that site places Hydrogen at High 21st of 78 and Argon 22nd of 78. Google's own model card says Hydrogen "did not show a material increase in general capability" over Argon, which is the same size of difference seen from the safety side.

Two more derived figures frame the rest. Google's "more than tripling" on Terminal-Bench 4.1 is 57 divided by 18.2 = 3.13 times, a 38.8 point difference. And of the 15 quantified performance claims in the launch post, 6 were measured by a named third party, 6 by Google, 1 by unnamed partners, and 2 by nobody the post names (derived count; the table in the claim ledger lists all 15).

Google sells Gemini access and Search, wrote the launch posts and the model card, and reports its own evaluations beside third party ones. Each claim below is attributed to whoever made it.

Read time. Google's four posts and the model card were saved by the lead between about 07:05 and 07:30 UTC on 13 November 2026; the methodology PDF, the Gemini API guide and pricing page, the Antigravity post, and the Artificial Analysis, Vals and CWE-bench pages were read by this site between 07:13 and 07:16 UTC the same morning. Scoreboards and prices move: everything below is as read then. Earlier briefings on this family, whose figures are not carried across: Argon, the Carbon announcement and the Carbon safety report.

The ladder: three Gemini 4 tiers, one index, different prices

Google's launch post describes Gemini 4 Carbon and Gemini 4 Argon as its frontier models for "deep, long-horizon reasoning", and Hydrogen as "built for rapid execution across diverse agentic workflows". The diagram places the three by what the saved pages print, with each figure marked as Google's or as an independent page's, and the gaps marked derived.

Zoom

The price gap is real but it is not a gap to every alternative. Hydrogen's list price is 25 percent below Argon's and Carbon's today (derived: 1.50 divided by 2.00 = 0.75, and 7.50 divided by 10.00 = 0.75), and 62.5 percent below the $4.00 and $20.00 that Google's pricing page says Argon and Carbon cost from 1 December 2026 (derived: 1.50 divided by 4.00 = 0.375). If Argon's token use on the Artificial Analysis tasks stayed the same, its cost per task there would double to about $9.94 after that date (derived: 4.97 times 2; this assumes the same tokens and cache mix, and no page prints it). But Google's own page lists Gemini 3.8 Flash at $0.75 and $3.75 through 31 December 2026 and $1.50 and $7.50 from 1 January 2027. A team moving from 3.8 Flash pays twice per token today (derived: 1.50 divided by 0.75 = 2.0) and the same per token from 2027, before any change in tokens used. The launch post's partner figure, 12 percent to 23 percent fewer tokens "compared to 3.8 Flash", is from unnamed partners in early testing.

xhigh is a price, not just a setting. Artificial Analysis lists 54 at xhigh against 51 at High: three more index points (derived) for $1.29 more per task (derived: 3.16 minus 1.87), which is 1.69 times the cost (derived), 1.73 times the time (derived: 13.7 divided by 7.9), and 2.76 times the output tokens (derived: 304M divided by 110M). Google's own Terminal-Bench 4.0 chart, read by eye, shows Hydrogen at about 57 percent at both High and xhigh, with xhigh costing more per task; the chart's numbers are not printed, so treat that reading as approximate. Google's guide says the same in words: xhigh "increases latency and token use and can cause overthinking on simpler prompts".

Zoom

Google's own chart, above, is the other half of the cost picture. It plots the same Artificial Analysis index against cost per task for Hydrogen, Claude Sonnet 5.5 and GPT-6.1 Terra. Read by eye, the GPT-6.1 Terra line ends near 51 at under one dollar a task, where Hydrogen at High reaches 51 at about $1.85, which matches Artificial Analysis's $1.87. The post's wording, "while often completing tasks faster than models in its class", is about time, and the companion time chart supports it. The cost chart shows that a cheaper tier than Argon is not the cheapest route to the same score on that index. Both readings are from Google's own figures and approximate.

The claim ledger: who measured each number in the launch post

The launch post carries many figures with no marker of who produced them. This table lists the 15 quantified performance claims in it (the safety figures follow in the next section), the party the saved pages say measured each, and what they print about the method. The count is this site's reading of those pages.

Zoom

The count: 6 third party (Artificial Analysis three times, Vals twice, arena.ai once), 6 Google, 1 unnamed partners, 2 not stated, which sums to 15 (derived). Third party measurement does not mean the third party chose the settings: the methodology PDF says "All of the results are run with the Gemini API with the highest thinking settings unless indicated otherwise below", while the headline index figure of 51 is at the default high level. The launch post and the methodology therefore mix settings in one announcement, which a reader has to track by hand.

Of the six third party claims, this site could read three against the pages: Artificial Analysis's index (51, 54, 50, matches), Vals Tax Agent Bench (52.30 percent, ranked first of 85, matches Google's 52.3), and the Vals Legal Research Bench (does not match; see the differences section). The Agent Arena rank (arena.ai) and the HLE-Diamond and AutomationBench-AA scores were not confirmed: arena.ai was not read, and the Artificial Analysis pages for those two benchmarks print no Hydrogen figure in the text this site saved.

The 57 and the 18.2. The post says Hydrogen is "more than tripling Terminal-Bench 4.1 scores (57% vs. 18.2%)" against Gemini 3.8 Flash. The methodology PDF says Hydrogen's 4.1 result is "self-computed", from a mini-swe-agent vision harness averaged over 5 attempts per task at high effort, and that "All results use cheating-adjusted accuracy". It does not say where 18.2 came from, how many tasks Terminal-Bench 4.1 has, or what the adjustment removes. The Terminal-Bench project's 4.0 announcement lists 4.1 as a future milestone for "verifier improvements (e.g. tamper-resistant verifiers)"; this site read only that announcement and no 4.1 leaderboard. Google's card also shows Claude Sonnet 5.5 at 63.9 percent on the same row, 6.9 points ahead of Hydrogen (derived: 63.9 minus 57.0).

The Orchestrator figures. The launch post says "On Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster". The Antigravity team's own post of the same date prints the identical sentence with "Terminal-Bench 4.1". No page gives the success rates behind "near-Carbon", the unit of "quota", or the baseline time.

What the model card says on safety and cyber, and what it prints

The model card is a 10 page PDF dated "November 2026" with no day. Its safety section opens with the sentence "The full Safety & Security Report for Gemini 4 Hydrogen will be available in the coming days." Counting that sentence and three "coming soon" pointers, four places in the card send the reader to a report the card describes as still to come (derived count, from the text layer).

What it says on frontier safety. Google evaluated Argon and Carbon under its Frontier Safety Framework (April 2026). For Hydrogen it ran a Material Capability Change assessment against Argon, "the most recent model version subject to a critical capability assessment", and says Hydrogen "did not show a material increase in general capability under our MCC assessment compared with Gemini 4 Argon and therefore a full suite of critical capability evaluations was not required". Because Google "could not rule out" that Argon and Carbon reached the biological CBRN CCL and the Cybersecurity CCL, it ran targeted assessments for Hydrogen in those two domains and "found that Gemini 4 Hydrogen does not reach the CBRN or cyber CCLs, though it reaches the CCL alert thresholds for both domains".

Of the card's seven frontier rows, 6 say "not reached" (4 carried forward from Argon, 2 after targeted testing), and 1, Loss of Control stealth and situational awareness, was not tested because there was no material change, so Google applies mitigations "as if the TCL had been reached" (derived counts; the rows are copied exactly under Technical details). Two rows, biological CBRN and cybersecurity, carry the alert threshold language.

The claim shape against the Carbon brief. This site's briefing on the Carbon safety report found a post and a report that read the model's cyber capability differently, with the report supplying seven cyber test rows to check. Hydrogen's pair reads differently: the launch post says nothing about cyber capability, and the card says Hydrogen showed no material increase in general capability against Argon, plus alert thresholds reached, with no rows at all. The same MCC method is used (no material change, so the full critical capability suite is skipped), but the evidence tables that made the Carbon finding checkable are, for now, absent. No Carbon figure is carried across.

The five safety rows that are printed. The card's "Training and Development Evaluation Results" are "automated evaluations and not human evaluation or red teaming", given as percentage point changes against Gemini 3.8 Flash. Three improved: text to text safety +0.80pp, multilingual safety +1.10pp, image to text safety +7.00pp. Two moved the wrong way by the card's own "higher is better" and "lower is better" notes: tone -0.65pp, and unjustified refusals +5.57pp (derived: 3 improved and 2 regressed, from the direction notes). The card says "unjustified refusals are higher, particularly for multimodal prompts". The launch post says "Hydrogen is a leap ahead" of 3.8 Flash on alignment and quotes a different measure, false refusals "under 0.4%" on HealthBench Pro. These are different metrics and neither page links them. None of the launch post's other safety figures (up to 81 percent less overreach, 12 times fewer misaligned actions, a 96.7 percent honesty rate on "our internal agent honesty evaluation", 99.9 percent chain of thought legibility over "more than 200,000 audited turns", 2.3 times on the Gray Swan prompt injection benchmark) appear in the card; they are Google's, and the baselines are not printed.

The card's known limitations are general: hallucinations, jailbreak resistance still being improved, "occasional slowness or timeout issues", and more tokens "especially at higher thinking levels", with a knowledge cutoff of March 2026.

Google's results table, and who ran each row

The card carries one results table as an image; the methodology PDF it cites repeats it. This site copied the rows and the values exactly. The Hydrogen column is Google's; the methodology says non Gemini numbers are "sourced from providers' self-reported numbers unless otherwise mentioned", so most of the other two columns are not Google's runs either.

Counting: Google ran Hydrogen's score on 7 of the 10 rows and took 3 from third party leaderboards (derived: 7 plus 3 = 10). Hydrogen is highest on 7 rows, 4 of them Google run (SWE-sweep, Agents' Last Exam, MKA, Chartography) and 3 third party (AutomationBench-AA, Vending-Bench 2, HLE-Diamond). It trails Claude Sonnet 5.5 on 3 rows, all of them Google run: Terminal-Bench 4.1 by 6.9 points, OSWorld-2.1 by 6.3 and Terminal-Bench-Science 0.1 by 8.5 (derived: 63.9 minus 57.0; 83.9 minus 77.6; 59.9 minus 51.4). That sits beside the launch post's sentence that Hydrogen "matches or outperforms competing models in its class across core developer and enterprise workflows". The post's phrase "in its class" is not defined, and the card's table does not say which of the two other models is in Hydrogen's class.

MKA is a benchmark Google published; the launch post calls it "new". The methodology says Google ran Hydrogen's score on it, and the other models' scores came from Artificial Analysis. The row is therefore a Google run on a Google benchmark with outside numbers beside it.

Zoom

Names, versions and two names for one thing

Full names are used throughout this briefing. The saved pages use short forms and, in several places, two names for one thing.

"Available to everyone today": what the pages say about tiers

The launch post's section heading says Hydrogen is "available to everyone today", and its text says "rolling out globally across our developer, enterprise, and consumer surfaces". The pages give tiers for each surface and no list of regions.

So "everyone" is true of Search and, for developers, of the free and paid API tiers; it is not true of Antigravity, where the Hydrogen model and the Orchestrator pairing sit behind paid plans, and the Carbon half of Orchestrator is Ultra only. The Search post says Hydrogen is "the biggest model upgrade to Search yet" and does not say which thinking level it uses, so the 51 and 54 on the index cannot be read onto Search results.

Where the sources differ

On CWE-bench. It is the independent cyber source for this launch and it ranks Hydrogen second of 14 models on v2, pass@1 68.1 percent at an average $2.63 a rollout with the Antigravity harness, against Argon at 65.0 percent and $3.70 (a 3.1 point gap, and Hydrogen's cost per rollout 29 percent lower, both derived). Its method is four rollouts per task on 250 tasks, 1,000 rollouts per model (derived: 4 times 250), "at high reasoning". The page does not say when Hydrogen's row was added, and Carbon is not on the v2 board. Collinear AI, which runs it, says on its own pages that "the training corpus is the product", so it too sells something near the benchmark. Google's launch post does not cite CWE-bench, and the model card prints no cyber score to compare it with.

What the pages state, and what they do not

Four labels doing the work of evidence

"Available to everyone". The post's own availability notes show Hydrogen sits behind paid plans in Antigravity and behind "auto model selection" for free Gemini app users. The label describes reach across surfaces, not equal access.

"Surpasses Argon". On the index Google cites it is one point at the default level, which Artificial Analysis lists. On the Vals Index it is below Argon, inside the printed error, and Google's own card calls the capability change from Argon not material. A tier can score higher on a composite and still not be a replacement for the tier above it.

"Priced flat". The flat refers to the price not changing across the 1 million token context window. It does not mean the price is flat over time or across tiers: the same pricing page lists a Priority tier at $2.70 and $13.50 per million tokens (1.8 times the standard rate, derived: 2.70 divided by 1.50), and Batch and Flex at $0.75 and $3.75. Argon and Carbon prices double on 1 December 2026, and 3.8 Flash reaches Hydrogen's price on 1 January 2027.

"Near-Carbon", "cheating-adjusted" and "fabricated claims". None is defined in the saved pages. Each carries a number or a comparison that looks measured. A defined term has a metric and a dataset; these have a label.

What to do, in order

Technical details

Everything in this section is printed by Google's pages and copied exactly: the model id, thinking level names, prices, the request shapes and migration lines from the Gemini API guide, the Decisions API model id from its post, and the model card's two tables. Google's pages print no log line, indicator, hash or detection rule, because the launch is not an incident; none is invented here. No step goes beyond what the pages print.

The model row below is the guide's own table row. It gives the id to put in the model field, the four thinking level names to put in thinking_level, and the paid tier price. A developer pins the id and the level together.

| Gemini 4 Hydrogen | `gemini-4-hydrogen` | `low`, `medium`, `high` (default), `xhigh` | $1.50 / 1M input, $7.50 / 1M output |

The quickstart request, as printed in the guide's Python tab. It sends a prompt to Hydrogen through the Interactions API and uses the default level, which the guide says is high with a 64K synchronous output limit. A developer runs it with an API key already set in the environment.

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-hydrogen",
    input="Build an interactive single-file HTML token bucket rate-limiter simulator with live sliders for burst size and request traffic."
)

print(interaction.output_text)

The same call with the thinking level set to the new top level. The generation_config field carries thinking_level. A developer sets this per request, and evaluates the lowest level that meets the quality bar.

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-hydrogen",
    input="Audit this payment retry queue for race conditions and rewrite the lock acquisition logic.",
    generation_config={
        "thinking_level": "xhigh"
    }
)

print(interaction.output_text)

The migration lines that change a request, as printed in the guide's checklists. A developer applies these when swapping from Gemini 3.8 Flash, Argon or Carbon; the last line is the one that changes behaviour silently if the level is left unset.

**Update model ID:** Change your model string from `gemini-3.8-flash` to `gemini-4-hydrogen`.
**Configure `thinking_level`:** Hydrogen defaults to `high` rather than `medium`, adds `xhigh`, and rejects `minimal` with an error.
**Move off explicit context caching:** Gemini 4 Hydrogen does not support explicit caching (`cachedContents` and `caches.create`).
**Remove sampling parameters:** Remove `temperature`, `top_p`, and `top_k` from requests. The server ignores custom values if sent.
**Check default thinking level:** Hydrogen defaults to `high` and supports `xhigh`, whereas Carbon and Argon default to `medium` and cap at `high`.

The pricing lines as the guide prints them. Cached input is a 95 percent discount (derived check: 1.50 times 0.05 = 0.075). Thinking tokens are billed as output. A developer budgets on output plus thinking tokens.

Input tokens: $1.50 per 1M tokens
Cached input tokens: $0.075 per 1M tokens via implicit caching (95% discount)
Output tokens, including thinking tokens: $7.50 per 1M tokens

The guide's advice on levels, as printed. A developer follows it as the vendor's migration advice; this site's suggestions are in the checklist above.

Start with `high` for most complex tasks, and reserve `xhigh` for hard, multi-step problems where `high` falls short.
Run your evaluations at all four levels with a high `max_output_tokens` so the limit doesn't cut answers short. Then pick the lowest `thinking_level` that meets your quality bar.
Using `xhigh` increases latency and token use and can cause overthinking on simpler prompts.

The Decisions API post, dated 11 November, prints a separate model id. It is the post Google links from Hydrogen's pages as a routing primitive. A developer who routes would meet this id; it is not Hydrogen.

client.decisions.create(
    model="gemini-decision-1",

The model card's training and development safety table, copied exactly. It reports automated evaluations as percentage point changes against Gemini 3.8 Flash. A defender or reviewer reads the direction notes: "higher is better" for the first four rows and "lower is better" for the last.

The model card's frontier safety table, copied exactly. It is the only place the card speaks about cyber capability. A reviewer reads the cyber row and the biological row for "CCL alert threshold reached", and notes the absence of any figure.

The question Google leaves open

Google's card says Hydrogen reaches the cyber alert threshold and that the evidence sits in a report due "in the coming days". Hydrogen is already the default model of the Antigravity agent in Gemini Managed Agents and answers in Google Search for all users. Which cyber tests produced the alert threshold finding, what did they score, and why does a model that reaches every surface on its launch day get its cyber numbers later?

Key facts