An iridescent processor sits beside two performance bars and a price tag.

Gemini 4 Hydrogen Pricing and Benchmarks: What Google’s Fast Model Costs and How It Scores

ByAIERA Blog Editorial Team

Key Takeaways

  • Google launched Gemini 4 Hydrogen on 12 November 2026, ending the leak phase covered in our earlier article: it is now a real, generally available model with a published price and benchmark table.
  • The price is $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1 million-token context window — double what Gemini 3.8 Flash costs today, but the same price Flash rises to on 1 January 2027.
  • Google reports Hydrogen scoring 51 on the Artificial Analysis Intelligence Index at its default high thinking level, ahead of Argon's 50, and 54 at its highest xhigh setting — but every benchmark figure is vendor-reported and none has been independently reproduced yet.
  • Against Gemini 3.8 Flash, Google reports Hydrogen finishing tasks 2.5 times faster, scoring 57% against 18.2% on Terminal-Bench 4.1, and cutting fabricated claims by more than five times.
  • Hydrogen is the fast execution tier of the Gemini 4 family, not a new frontier model: Carbon and Argon handle deep reasoning, while Hydrogen does the everyday coding and agentic work, with a new Orchestrator in Google Antigravity routing between them.
  • The Cloud model page adds constraints the announcement does not: Hydrogen is a short-term generally available release with a minimum 45-day retirement notice, and it has no RAG Engine, tuning or Live API support.

The Leaks Are Over: Hydrogen Is a Real Product Now

On 10 November 2026, this site published an article explaining what the leaks said about Gemini 4 Hydrogen. That article was careful, because it had to be: Google had announced nothing, no Hydrogen model ID appeared in any Google product, and every claim traced back to posts on X and tester accounts of blind arena battles. Two days later, the story changed completely. On 12 November 2026, Google DeepMind announced Gemini 4 Hydrogen as a real, generally available model, with a price, a benchmark table, a model card and a global rollout.

This article covers only the new material: the launch itself, the pricing, the benchmarks Google published, and how Hydrogen compares with Argon, Carbon and the model it replaces, Gemini 3.8 Flash. It does not repeat the leak history. If you want the rumour trail — the codename's origin, the arena sightings under the "gemini-3.8-flash" label, and the unconfirmed speed claims — that is in the earlier piece. What follows is what Google has now said on the record, plus the limits and caveats that the announcement itself does not spell out.

The announcement came from Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, in a post titled "Introducing Gemini 4 Hydrogen for Agentic Workflows", published on 12 November 2026. His framing is worth quoting because it sets the whole positioning of the model: while the frontier models, Gemini 4 Carbon and Gemini 4 Argon, handle deep reasoning, Hydrogen is built "for rapid execution across diverse agentic workflows" — writing code, running tests, verifying changes, doing research, and carrying multi-step tasks to completion. Google calls it, in coverage of the launch, its daily driver for agentic workflows. That is a deliberately unglamorous job description, and it is the point: Hydrogen is not the smartest Gemini 4 model. It is the fast one.

The launch also quietly settled two questions the leaks had left open. The first is the name: Google's three posts on 12 November never use the name "Gemini 4 Flash". They describe Hydrogen as the upgrade for teams moving from Gemini 3.8 Flash, which is the place the leaks gave it, but the product name is Hydrogen, not Flash. The second is the date: two X accounts had estimated a release around 18 November. The actual launch was six days earlier, on 12 November — a reminder that leak timelines are guesses, even when the underlying story turns out to be right.

The Price: $1.50 and $7.50, Flat Across 1 Million Tokens

The pricing is the single most concrete fact in the announcement, and Google stated it with unusual precision. In the announcement post, Kavukcuoglu wrote: "Gemini 4 Hydrogen 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."

Three things about that sentence matter. First, the numbers themselves: $1.50 per million input tokens and $7.50 per million output tokens. Second, the word "flat": the price does not rise as the context window fills, so a prompt that uses 900,000 tokens costs the same per token as one that uses 9,000. Third, the window: the full 1 million-token context is included at that price, with input, thinking and output tokens sharing the same 1,048,576-token window, according to the model documentation read by launch coverage.

That flat structure is a deliberate contrast with how some models price long context. Several providers charge more per token once a prompt passes a threshold, which makes long-context work hard to budget. Hydrogen's flat price makes cost modelling simple: the bill is a function of tokens in and tokens out, nothing else. For teams running agents that accumulate long histories — a coding agent that keeps a whole repository in context, or a research agent that synthesises dense PDFs and spreadsheets — that predictability is arguably as important as the headline rate.

The price also needs to be read against the model Hydrogen replaces. Gemini 3.8 Flash, the current Flash model, costs $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026. Hydrogen is exactly double that. But the comparison flips on 1 January 2027, when Gemini 3.8 Flash rises to $1.50 and $7.50 — the same price as Hydrogen. From that date, the old Flash model and the new one cost the same, and the only reason to stay on 3.8 Flash would be inertia. Google's pricing page for AI Studio, as read by launch coverage, also lists cached input for Hydrogen at $0.075 per million tokens, which matters for agents that re-read the same context repeatedly.

Hydrogen against Argon and Carbon

The more interesting comparison is inside the Gemini 4 family, because that is where Hydrogen's position becomes clear. Gemini 4 Argon, the confirmed frontier model that reached general availability on 13 October 2026, costs $2 per million input tokens and $10 per million output tokens through 30 November 2026, rising to $4 and $20 from 1 December 2026. So Hydrogen, at $1.50 and $7.50, is cheaper than Argon on both counts — 25% cheaper on input and 25% cheaper on output at Argon's current price, and far cheaper once Argon's December price rise takes effect.

Gemini 4 Carbon, the strongest model in the family, has no published price at all. Google announced Carbon on 28 October 2026 for cyber defenders only, with access phased through its Fairwind Program, and it has no listing on the Gemini API pricing page. That tells its own story: Carbon is not a product you can buy yet, while Hydrogen is a product you can call today. The family, as it stands on 13 November 2026, is a three-tier structure with a clear price ladder: Hydrogen at $1.50/$7.50 for fast execution, Argon at $2/$10 (rising to $4/$20) for frontier reasoning, and Carbon at an unpublished price for the strongest, defender-only work.

Against the wider market, Hydrogen sits in the middle of the mid-tier. Launch coverage comparing token prices put Claude Haiku 4.5 at $1.00/$5.00, Claude Sonnet 5 at $2.00/$10.00, Grok 4.6 at $2.00/$6.00, Mistral Large 3 at $0.50/$1.50 and DeepSeek V4 Pro at $1.32/$3.96. Hydrogen is more expensive than Haiku 4.5 and DeepSeek V4 Pro, cheaper than Sonnet 5 and Argon, and roughly in line with the fast-tier pricing of its class. Price alone will not decide whether teams switch; the benchmark and speed claims will matter at least as much, and those need the same careful reading.

The Benchmarks: What Google Claims, and Why They Are All Vendor-Reported

Google published a broad benchmark table with the Hydrogen announcement, and the headline numbers are striking. The single most important thing to understand about them is that they are all vendor-reported: Google measuring Google's own model, on benchmarks Google chose, using Google's own harness. None of them had been independently reproduced at the time of writing. That does not make them false. It makes them a starting point for verification, not a conclusion.

The headline claim is the Artificial Analysis Intelligence Index, an independent index that Google cites in its own announcement. At its default high thinking level, Google says Hydrogen scores 51, ahead of Argon's 50. At its highest thinking setting, xhigh, it reaches 54. That is a narrow lead over Argon at default — one point — and a wider one at xhigh. The thinness of the default margin matters: a one-point edge on a composite index is real but small, and it is measured at Hydrogen's default configuration against Argon's presumably default configuration. It is also worth noting what the index measures: general intelligence across a basket of tasks, not speed or agentic reliability. A model can lead on that index and still be the wrong choice for a workload that depends on tool calling or long-horizon execution.

The speed claims are where Hydrogen's case is made most forcefully. Google reports that, at its high default setting, Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash. On Terminal-Bench 4.1, a benchmark for agentic terminal work, Google reports Hydrogen scoring 57% against Flash's 18.2% — more than tripling the result — and cutting fabricated claims by more than five times. In early partner testing across graph, SQL and data-heavy enterprise workloads, Google reports Hydrogen running 2 to 2.8 times faster than 3.8 Flash while using 12% to 23% fewer tokens. Google also says Hydrogen recovers cleanly when a tool call fails or a search returns empty, and that it natively batches file-read and search calls in parallel.

Those are dramatic numbers, and the Terminal-Bench jump is the one that will circulate most. But it is also a jump over 3.8 Flash, a model from the previous generation, on a benchmark designed for exactly the agentic terminal work Hydrogen is built for. A tripling on a benchmark that plays to the new model's strengths is impressive; it is not the same as a tripling across all tasks. The honest reading is that Hydrogen looks like a large step forward for agentic terminal work specifically, and that the size of the step on other workloads is still to be established by independent testing.

The agentic benchmark table

Beyond Terminal-Bench, Google's announcement lists a set of agentic and professional-work benchmarks. Hydrogen ranks fifth on Agent Arena's live ranking. It leads its class on HLE-Diamond at 58.8%. It ranks first overall on the Vals Tax Agent Bench at 52.3%, and ties for first on the Vals Legal Research Bench at 55.8%. It leads its class on AutomationBench-AA. For multimodal knowledge work, Google reports 53.4% on the Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1, with a native 1 million-token input window and up to 1 million output tokens for synthesising dense PDFs, spreadsheets, charts, audio and recorded video into cited reports. Google also says Hydrogen outperforms similar-tier models on SWE-sweep, a software-engineering benchmark.

There is one piece of independent corroboration in this list, and it is worth singling out. Vals AI, an independent evaluator, ranks Hydrogen first of 85 models on its Tax Agent Bench at 52.3% when run at xhigh reasoning, and sixth of 37 on the Vals Index at 68.49%, according to launch coverage that read Vals AI's tables. That is a genuinely independent result, not a Google self-measurement, and it supports the claim that Hydrogen is strong at professional agentic work. But it is one evaluator, on one class of task, at the highest reasoning setting — which is also the most expensive and slowest setting. It is a promising signal, not a full verification.

The safety figures in the announcement deserve the same treatment. Google reports Hydrogen reducing overreach in agentic tasks by up to 81% and under-service by up to 79%, cutting misaligned actions and task shortcutting by 12 times, with 96.7% honesty, 99.9% chain-of-thought legibility, and over 2.3 times stronger resistance on the Gray Swan IPI benchmark. These are Google DeepMind's own internal evaluations. They point in a welcome direction — a fast model that is also more careful about when to act and when to ask — but they are self-assessments, and the "up to" in the overreach and under-service figures means the best case, not the typical case. Teams deploying agents unsupervised should reproduce these on their own traces before relying on the margins.

Four Thinking Levels, and Why They Change the Price Math

Hydrogen ships with four configurable thinking levels: low, medium, high as the default, and a new xhigh tier. Thinking levels control how much reasoning the model does before answering, and they are the mechanism behind the benchmark spread: 51 on the Artificial Analysis index at high, 54 at xhigh.

The levels matter for cost as well as quality. A higher thinking level means more thinking tokens, and thinking tokens are billed as output tokens at $7.50 per million. The xhigh setting that produces the 54 index score and the first-place Vals ranking is therefore the most expensive way to run the model, and also the slowest. Launch coverage that read the Artificial Analysis index tables put the estimated cost per task at $1.87 for Hydrogen at high, taking about 6 minutes, against $3.16 at xhigh, taking about 13 minutes. Those are estimates from one evaluator's tables, not Google's figures, but they illustrate the trade-off: the best benchmark numbers come at roughly 70% more cost and more than double the time.

For everyday use, that means the headline benchmark scores and the headline price describe different configurations. The $1.50/$7.50 price is the rate; what a task actually costs depends on the thinking level, and the thinking level that produces the best scores is not the default. Teams should test Hydrogen at the level they will actually run in production, not at the level that produces the best press-release number.

Orchestrator and Antigravity: Hydrogen as Half of a Pair

The most interesting architectural claim in the announcement is not about Hydrogen alone, but about how it works with Carbon. Google introduced Orchestrator in Google Antigravity, its agentic coding platform, which dynamically switches between models like Carbon and Hydrogen within the same session. The idea is a division of labour: Carbon handles deep architectural judgement, while Hydrogen powers worker and verifier agents in isolated workspaces, doing the disciplined execution — writing the change, running the test, reading the failure, trying again.

Google reports that, on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while consuming 24% less quota and completing tasks 19% faster on average. That is a system-level result, not a model-level one, and it depends entirely on routing accuracy: if the orchestrator sends low-level tasks to Carbon or deep-debugging tasks to Hydrogen, the cost and latency upside disappears. But if the routing works, the implication is significant. It means the interesting product is not Hydrogen alone, but the pairing — a frontier model for judgement and a fast model for execution, with the routing between them handled by the platform rather than by the developer.

This is also where the "anti-gravity angle" from the leak coverage finds its answer. The leaks had suggested a `gemini-3.8-flash-router` string in Antigravity metadata might be connected to the new Flash model. The announcement confirms the shape of that guess: Antigravity is where the multi-model routing lives, and Hydrogen is the execution tier it routes to. What the leaks could not establish — and what the announcement now confirms — is that the routing is a named product, Orchestrator, with published (if vendor-reported) performance figures.

Where Hydrogen Is Available, and Who Gets It Free

Hydrogen rolled out on 12 November across all Google AI Studio tiers, the Gemini API, Gemini for business, Google Search for all users, the Gemini app, and paid Google AI Pro and Ultra subscribers in Google Antigravity. Gemini in Google Sheets starts with eligible business subscriptions and AI Pro and Ultra subscribers.

Two details in the rollout are easy to miss and both matter. First, Hydrogen is available on the Gemini API free tier, rate-limited, unlike Carbon, which is paid-only. That makes Hydrogen testable without a billing relationship, which is how most developers will first meet it. Second, free Gemini app users do not get to choose Hydrogen directly: the announcement footnote says they access it through auto model selection only, with Google's routing deciding when it is used. Paid Pro and Ultra subscribers can use it in Antigravity directly.

In Search, Google says Hydrogen better parses multi-part, typo-heavy and highly specific queries, and widens query fan-out so users see links they might previously not have discovered. That is a consumer-facing change rather than a developer one, but it is the surface where most people will encounter Hydrogen first — not as a model they choose, but as a better answer in Google Search.

The model ID is `gemini-4-hydrogen`, the context window is 1,048,576 tokens shared across input, thinking and output, synchronous output is capped at 65,536 tokens per call with longer output requiring background mode, and the knowledge cutoff is March 2026, according to the model card and documentation read by launch coverage.

The Constraints the Announcement Does Not Mention

The announcement is a launch post, and launch posts do not lead with limitations. The Google Cloud model page, read by coverage after the launch, adds four constraints that change the buying decision.

The first is the most important: Hydrogen is a short-term generally available release, subject to a minimum 45-day retirement notice policy. That means Google can retire the model with as little as 45 days' notice. For a team building a product on Hydrogen, that is a real planning constraint — shorter than the runway most teams want for a core dependency — and it is the number a buyer should place beside the price.

The second is sampling: the documented defaults are temperature 1.0, top_p 0.95 and top_k 64, all marked fixed, with candidate count fixed at 1. There is limited room to tune the model's sampling behaviour.

The third is caching: Hydrogen supports implicit caching only, with a 4,096-token minimum, and durable caching enabled by default. Explicit context caching is not supported.

The fourth is the missing features: the Cloud model page lists no RAG Engine, tuning or Live API support for Hydrogen. Teams that need those capabilities will have to use another model or wait.

None of these constraints makes Hydrogen a bad model. They make it a model with a specific shape: fast, cheap per token, broadly available, but short-term, fixed in its sampling, and without some of the platform features that longer-lived models carry. That shape suits experimentation, agentic workloads and cost-sensitive production use. It suits a core, long-lived dependency less well, unless Google extends the commitment.

What We Could Not Verify

Several things remain unverified, and they should be stated plainly rather than filled in with guesses.

No independent speed measurement of Hydrogen has been published in the sources reviewed for this article. Google published no tokens-per-second figure in its announcement — the rumoured 300 to 500 tokens per second from the leak coverage remains a rumour, and the official numbers are task-completion multiples, not throughput figures. Until an independent evaluator publishes a speed measurement with a stated prompt, output length and method, Hydrogen's speed should be described as "Google reports 2.5 times faster than 3.8 Flash on its tasks", not as a throughput number.

No independent benchmark of Hydrogen's general intelligence has been published either. The Artificial Analysis Intelligence Index score of 51 is cited by Google, not published by Artificial Analysis as its own measurement of the model. The Vals AI ranking is the one independent result found so far, and it covers professional agent work at the highest reasoning setting, not general capability.

The 45-day retirement notice, the fixed sampling parameters, the caching limits and the missing RAG, tuning and Live API support are reported from the Cloud model page by launch coverage, not read directly from Google's documentation for this article. They should be confirmed against the live model page before being relied on for a procurement decision.

Finally, the launch does not answer the question the leaks raised about what happens to Gemini 3.8 Flash. Google has said nothing in the announcement about retiring Flash, and the price rise to $1.50/$7.50 on 1 January 2027 is the only signal so far. Whether Flash is retired, kept as a legacy option, or quietly wound down remains unannounced.

The Bottom Line: A Repricing and Latency Play, Not a New Ceiling

Hydrogen is best understood as a repricing and latency play, not a new reasoning ceiling. Google has taken the Gemini 4 family, which had a frontier model in Argon and a stronger, restricted model in Carbon, and added the tier it was missing: a fast, affordable model for the everyday work that agents actually do all day. The price is clear, the benchmarks are strong on agentic work, and the Orchestrator pairing with Carbon is the most interesting idea in the launch.

But the numbers that will circulate — 2.5 times faster, 57% against 18.2%, 51 against 50 — are all Google's numbers, on Google's benchmarks, measured Google's way. The one independent result so far, from Vals AI, is encouraging. Everything else awaits independent reproduction. The sensible response is to test Hydrogen on your own workload, at the thinking level you will actually run, and to treat the vendor benchmark table as a hypothesis to validate — while keeping an eye on the 45-day retirement notice, which is the one number in this launch that is not a claim about performance at all, but a constraint on planning.

For background on the Gemini family, see our guide to Google Gemini 3: release date, features and what is new, and for the model Hydrogen sits below in the family, see our guide to Gemini 4 Argon pricing and availability. Our explainer on what AI agents are and how they work sets out why the agentic benchmarks in this launch matter more than the general intelligence scores.

Sources and Further Reading

  • Google DeepMind, "Introducing Gemini 4 Hydrogen for Agentic Workflows", 12 November 2026: https://blog.google/intl/en-in/products/introducing-gemini-4-hydrogen/
  • AI Mastery, "Google Claims Gemini 4 Hydrogen Runs 2.5x Faster Than 3.8 Flash": https://www.aimastery.page/news/gemini-4-hydrogen-fast-execution-tier
  • DEV Community, "Gemini 4 Hydrogen: Google's Fast Execution Model Beats Argon on One Benchmark — Here's the Catch": https://dev.to/tng_thnhnhn_c933ee88b/gemini-4-hydrogen-googles-fast-execution-model-beats-argon-on-one-benchmark-heres-the-catch-2ch6
  • TBN Grow, "Gemini 4 Hydrogen pricing at $1.50: time to drop Gemini 3.8 Flash?": https://tbngrow.com/news/gemini-4-hydrogen-gia-api-thay-flash-2026-11-13?lang=en
  • My Industry News, "Gemini 4 Hydrogen Is Cheap per Token. The 45-Day Retirement Notice Is the Real Constraint.": https://myindustrynews.net/TenantSites/1bf91c5e-dbc9-4e91-a10c-d949ca1d094d/2026/11/68abb8cc-c65d-4946-88cd-0c7ac8e4d068.html

Similar Posts