Gemini 4 Hydrogen: Google's Flash Replacement Drops the Flash Name and Argues About Finishing

The claim G​oogle DeepMind put in the announcement is narrow and specific: G​emini 4 Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash and beats Gemini 4 Argon on the Artificial Analysis Intelligence Index. Both halves are checkable, and the second one is true — on the independent harness Hydrogen scores 54 against Argon's 50. The first half is where the launch gets interesting, because the same independent harness puts Hydrogen at 13.5 minutes per index task and Gemini 3.8 Flash at 14.8, which is a 10% difference and not a 150% one. G​emini 4 Hydrogen is real, it shipped generally available on 12 November 2026 at $1.50 per million input tokens and $7.50 per million output, and it is the first model in the throughput slot to be sold under a codename rather than a Flash label. What it is not is a simple speed upgrade over the model it replaces.

What actually shipped on 12 November

G​oogle's model card is dated the same day and is unusually explicit about lineage: Hydrogen is adapted from Gemini 4 Carbon, the frontier model in the same family. It is not a smaller model trained separately. It is the same base with a different post-training and serving profile, which is why the capability gap between it and its siblings is smaller than the price gap suggests.

• Context — one 1,048,576-token window shared across input and output, with 64K tokens of default synchronous output and up to the full million through background long decode.

• Inputs and outputs — text, images, audio, video and PDFs in; text out. No image generation, no audio generation, no Live API.

• Thinking levels — configurable, and Artificial Analysis publishes its headline score at the xhigh setting, which is not the default. G​oogle's own comparison numbers sit at the default high.

• Knowledge cutoff — March 2026, stated plainly in the model card's limitations section, with the caveat that some domains are less current. That single line resolves a question the earlier G​emini 4 models left open.

• Distribution — G​emini app, G​emini API, G​oogle AI Studio, G​oogle Sheets, G​emini for business, G​oogle Antigravity and G​oogle Search, with a free tier on the API. No gated access phase.

• Safety reporting — a separate Safety and Security Report is promised "in the coming days" and does not exist yet, so everything below is self-reported.

The 2.5x claim, and what the outside harness measured

Two numbers are doing the work in G​oogle's announcement, and they answer different questions. The first is throughput: 190 output tokens per second measured by Artificial Analysis, against 112 for Gemini 3.8 Flash, 65 for Gemini 4 Argon and 72 for Gemini 4 Carbon. That is a genuine 1.7x on raw generation speed and roughly three times either G​emini 4 sibling.

The second is time per completed task, and here the picture flips. Artificial Analysis finishes an Intelligence Index task with Hydrogen in 13.5 minutes, with Gemini 3.8 Flash in 14.8 minutes, with Carbon in 28.4 and with Argon in 30.0. Against Argon that is a 2.2x advantage. Against the model G​oogle named in the launch post it is 1.1x. If the 2.5x figure came from a specific internal task set, it is a claim about that task set — the general-purpose outside number is close enough to parity that no one outside G​oogle has reproduced the gap.

The likely explanation is verbosity, and it cuts both ways. Hydrogen generated 304 million output tokens to complete the index, against 72.8 million for the median model and 126 million for Gemini 3.8 Flash. It is producing more than twice the text per task and still finishing slightly sooner, because it produces that text fast. Speed and verbosity are the same fact seen from two ends.

What G​oogle's own table says, with the caveats attached

Every figure in this section is vendor-reported, from the 12 November model card, and none of it has been reproduced outside G​oogle. The comparison set is Claude Sonnet 5.5 and GPT-6.1 Terra, not the Flash model being replaced.

• SWE-sweep — 13.5% for Hydrogen against 3.1% for Claude Sonnet 5.5 and 5.8% for GPT-6.1 Terra, the largest relative gap in the table.

• Terminal-Bench 4.1 — 57.0% for Hydrogen against 63.9% for Claude Sonnet 5.5, with no score published for GPT-6.1 Terra. This is the one agentic coding benchmark where Hydrogen loses, and G​oogle labels it cheating-adjusted.

• Knowledge work — Agents' Last Exam v1.1 at 71.6% (60.4%, 69.8%), AutomationBench-AA at 76.1% (71.8%, 60.2%), and Vending-Bench 2 at a $14,549 bank balance against $11,165 and $9,107.

• Computer use — OSWorld-2.1 at 77.6% against 83.9% for Claude Sonnet 5.5 and 74.2% for GPT-6.1 Terra. A second loss to the same competitor.

• Multimodal and science — MKA 53.4%, Chartography 66.6%, Terminal-Bench-Science 0.1 at 51.4%, and HLE-Diamond at 58.8% against 53.3% and 47.3%.

Read the losses together and a shape appears: Hydrogen is strong where the job is to accumulate work over many steps and weaker where the job is to hold a long tool-using session together. That is consistent with a model tuned for throughput rather than endurance, and it is the difference that matters if you are choosing between it and a frontier model for an agent loop that runs for an hour.

The independent numbers are the ones to budget against

Artificial Analysis is the only instrument here that did not come from the vendor, and its cost-per-task figure is the most useful single number in the launch. Hydrogen averages $3.16 of inference to complete one index task at xhigh, and $1.87 at the default high. Gemini 4 Argon costs $4.97, Gemini 4 Carbon $7.76, and Gemini 3.8 Flash — the cheap model — costs $7.77, because it burns 126 million tokens at a low price to get there.

That inversion is the actual argument for Hydrogen and it is stronger than the speed claim. The model that costs a quarter as much per token as Carbon costs a fifth as much per finished task, and less than a quarter of what the model it replaces costs. Per-token pricing would have told you the opposite.

Two caveats belong with it. The first is that cost per task is a property of the index's task mix, not of your workload; a pipeline that emits 200-token answers and stops will see the per-token price and none of the per-task saving. The second is latency: Hydrogen needs 37.8 seconds to its first answer token at xhigh — 16.7 at the default high — against 17.3 for Gemini 3.8 Flash. At the setting the launch coverage quoted it is fast at generating and slow at starting, so an interactive surface will feel slower than the model it replaced even though the agent loop finishes sooner.

Safety reporting that includes a regression

The model card carries an unusual admission for a launch document. Against Gemini 3.8 Flash, automated content-safety scores improved on text (+0.80 percentage points), multilingual (+1.10) and image-to-text (+7.00), and tone slipped by 0.65. Unjustified refusals — the model declining benign requests — got worse by 5.57 points, concentrated in multimodal prompts.

On frontier safety, G​oogle ran a Material Capability Change assessment against Gemini 4 Argon and concluded Hydrogen shows no material increase in general capability, so the full critical-capability suite was not required. Targeted CBRN and cyber testing found the model does not reach either critical capability level, though it does reach the alert threshold for both. Those are self-assessments with the underlying report still listed as forthcoming.

Where this leaves the Flash tier

G​oogle has not retired Gemini 3.8 Flash. It is still in the catalogue, still $0.75 / $3.75 per million tokens, and still the right answer for short, latency-sensitive calls: at 17.3 seconds to first answer token it starts twice as fast as Hydrogen at xhigh, and roughly level with Hydrogen at its default. What changed is the default: the model G​oogle now ships to free-tier API users and to the Antigravity agent in G​emini Managed Agents is the one with the codename, and the tier label that used to describe that slot is gone.

For teams running a mixed stack, the practical consequence is that the two models want different jobs. Hydrogen for execution — tool calls, file edits, test runs, parallel subagent turns — and a frontier model for the planning turn. That is exactly the arrangement G​oogle documents, where Carbon plans interfaces and resolves hard roadblocks and Hydrogen takes the code edits, and it is the arrangement OrcaRouter is built to hold: Gemini 4 Carbon, Gemini 4 Argon, Gemini 3.8 Flash and Gemini 3.1 Pro Preview are all callable from one key at provider list price with no markup, so the 1 December repricing of Carbon and Argon lands here the same day it lands on G​oogle's card, and a rate-limited free tier can fall back to a paid path without a second contract. G​emini 4 Hydrogen is not in our catalogue — it comes from G​oogle's API and AI Studio — so the honest configuration is Hydrogen called direct for execution with the routed models behind it for planning and failover.

What to watch instead of another benchmark

The date that will tell you more than any score is 1 December 2026. Gemini 4 Carbon and Gemini 4 Argon both double on that date, from $2.00 / $10.00 to $4.00 / $20.00 per million tokens. Hydrogen's card carries no dated step at all. If that asymmetry holds, G​oogle has created a stable two-tier family — a $4/$20 ceiling and a $1.50/$7.50 workhorse — and the Flash name is unlikely to return. If Hydrogen moves onto the same schedule, then what shipped this week was an introductory rate with a codename attached.

The second thing to watch is the safety report, which the model card says is coming and which is the only document that could contradict the capability assessment above. The third is whether anyone outside G​oogle reproduces a 2.5x task-completion gap. Until one does, the defensible version of the claim is the one the independent harness supports: Hydrogen is the cheapest G​emini 4 per completed task, roughly three times the throughput of its siblings, slow to start and extremely verbose — and the model it replaced is now the budget option for short calls rather than the default for everything.