The Fast Model Takes the Default Slot: Google Launches Gemini 4 Hydrogen, Four Thinking Levels, 2.5x Faster Than 3.8 Flash
On November 12, Google launched Gemini 4 Hydrogen, a fast execution model for everyday coding and agentic workflows, with four thinking levels — low, medium, high and xhigh — flat pricing of $1.50 and $7.50 across a 1-million-token window, 51 on the Artificial Analysis index at the default level against 50 for Argon, 54 at xhigh, and an Antigravity Orchestrator that switches between Carbon and Hydrogen automatically.

The flagship thinks; the fast model does the work
On November 12, Google introduced Gemini 4 Hydrogen on its official blog, in a post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. The timing is worth noting: just three days earlier, Gemini 4 Carbon had gone generally available in GitHub Copilot, and Gemini 4 Argon had landed not long before that. Three Gemini 4 models in about a month makes Google's intent clear. It is not collecting one smartest model; it is assembling a team with a division of labour. Carbon and Argon take on the hardest long-horizon reasoning, while Hydrogen does the large volume of everyday work, fast and cheap. In Google's own words, Hydrogen is a fast model built for everyday coding and agentic workflows, the company's "daily driver".
What matters most about this launch is not any single score but a change of role. For the past year, model launches have been about being smarter: bigger context windows, higher benchmark scores, harder reasoning tests. Hydrogen goes the other way. Its pitch words are speed, efficiency and price. Google says it finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking setting, more than triples Terminal-Bench 4.1 from 18.2% to 57%, and cuts fabricated claims by more than five times. For the first time, the fast model is not just "good enough" but is being placed in the default slot. For developers whose agents run hundreds of loops a day, that matters more than a few more points on a flagship benchmark.
The price is the hardest fact in the announcement. Hydrogen is priced flat across its entire 1-million-token context window: $1.50 per million input tokens and $7.50 per million output tokens. Note the word flat: input, thinking and output all cost the same rate, however full the context gets, with no length-based tiers. Google also gives it four thinking levels: low, medium, high (the default) and a new xhigh tier. The higher the level, the deeper, slower and more expensive the reasoning; the lower the level, the faster and cheaper the answer. A small edit gets low; a stubborn debugging session gets xhigh. The bill follows the difficulty of the task instead of being one flat rate for everything.
Four thinking levels: how to choose
The four thinking levels are the most product-minded design in this release. Older models made you pick a "fast version or a smart version", or hid reasoning depth inside a black box. Hydrogen turns it into an explicit dial. Google describes the levels as a "smooth, predictable progression" so you only wait and pay for the reasoning depth a task requires. That sounds like marketing, but it maps onto a real pain: 90% of calls in an agent loop need no deep thought at all — reading a file, changing one line of config, judging the result of a tool call. Running those on a flagship model is like hiring a head chef to chop spring onions.
The levels are tied to scores. At the default high level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, one point above Argon's 50; at xhigh it reaches 54. Let us be honest: 51 versus 50 is not worth boasting about. A one-point gap on that index is within noise and version churn. But it signals something: Google is willing to put a "fast model" on the same scoreboard as its flagships, and it did not lose. That is the real signal — the line between fast and smart is blurring. Google keeps one caveat in reserve: xhigh is a new tier, and it adds latency and token use. The blog gives no absolute latency figures, only relative multiples, and we will come back to that.
For people who vibe-code, the practical meaning is a change in cost structure. Until now, picking a model for an agent loop was mostly binary: a cheap Flash tier for volume, a flagship for the hard parts. Hydrogen fills the middle: everyday execution runs on high, a hard nut gets a single switch to xhigh, and architecture judgement goes to Carbon. Google explicitly recommends this pairing in the blog: Carbon and Argon for deep architectural judgement, Hydrogen as the execution engine. That is not a pleasantry; the multi-model collaboration described below is designed around exactly this split.
Against 3.8 Flash: where it is actually better
Hydrogen's direct predecessor is Gemini 3.8 Flash, and Google frames it as a major standalone upgrade, with four sets of comparison numbers. First, speed: tasks finish 2.5 times faster at the default high setting. Second, Terminal-Bench 4.1: 57% versus 18.2%, more than triple. Third, fabricated claims: down by more than five times. Fourth, from early partner testing on complex graphs, SQL and data-heavy enterprise workflows: 2x to 2.8x faster while using 12% to 23% fewer tokens than 3.8 Flash. Read the four groups separately: the first three are Google's own benchmark claims, the fourth is early partner testing, and none of them is an independent third-party replication. The direction is consistent; the precision deserves a discount.
Another set of numbers deserves its own paragraph. On iterative repository coding and debugging, Google says Hydrogen outperforms similar-tier models on SWE-sweep, with "focused, surgical edits" that fix bugs without rewriting unrelated files, producing cleaner code that is closer to production-ready. More important is the method: it natively batches file-read and search calls in parallel instead of inspecting files one by one. For a large repository, that is where the feeling of speed really comes from: a model that reads ten files at once feels very different from one that reads one at a time, whatever the parameter count. On Agent Arena's live benchmark Hydrogen ranks fifth. It leads its class on HLE-Diamond at 58.8%, ranks first overall on Vals Tax Agent Bench at 52.3%, ties for first on Vals Legal Research Bench at 55.8%, and leads its class on AutomationBench-AA. The benchmark names are technical; the conclusion fits in one sentence: Hydrogen is not only a fast coding model, it can also hold its own on agent tasks that need research and judgement.
Multimodal work is a second thread. Hydrogen pairs a native 1-million-token input with up to 1 million output tokens, synthesising dense PDFs, spreadsheets, charts, audio and recorded video into polished, cited reports. It scores 53.4% on the new Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1, a benchmark for end-to-end professional workflows. On web development, Google says it generates more functional, feature-complete layouts in fewer prompts; the blog's example is synthesising real-time global lightning strike data into an interactive 3D globe. All of these claims come from Google's own testing, and that premise should stay attached to them.
The genuinely new part: Orchestrator lets models take turns
If Hydrogen alone were the story, this would be a routine model refresh. The part worth its own section is Orchestrator, launched alongside it. This is a new feature in Google Antigravity: instead of running one model from start to finish, it dynamically switches between models inside the same session — Carbon for judgement, Hydrogen for execution — with the system deciding when to hand over. On Terminal-Bench 4.0, Google reports that Orchestrator reached near-Carbon task success rates while consuming 24% less quota on average and finishing 19% faster. If that holds up, it matters more than Hydrogen's own leaderboard: a "fast plus smart" combination can get close to smart-only results while costing a quarter less and finishing a fifth faster.
Parallel agents follow the same logic: Carbon acts as coordinator while Hydrogen powers worker and verifier agents, building and testing changes simultaneously across isolated workspaces. This has a direct lesson for vibe coding workflows. Until now, one person directed one agent; now it is a team of agents — the expensive one decides, the cheap ones do the work, check it and redo it. The bulk of the cost was never the hardest step; it is the hundreds of executions and verifications. Hydrogen's job is to bring down the unit price of execution while keeping the error rate low enough that rework does not eat the savings. Rework, after all, usually costs more than the model call itself.
Orchestrator has a gate, though: it currently lives only in Google Antigravity, and Hydrogen inside Antigravity is available only to paid Google AI Pro and Ultra subscribers. Free users and AI Plus subscribers remain on 3.5 Flash-Lite for now. In other words, the cheapest "automatic handover" combination currently sits behind a paywall. That is Google's usual pattern: paid users fund the new workflow first, and it trickles down later.
Where ordinary users will meet it: Search and the Gemini app
Hydrogen is not only a developer model; on the same day it went into Google Search and the Gemini app for all users. Search changes in three ways: it understands the true intent of a question better — multi-part queries, hard-to-read typos, very specific requests; responses become more visual and easier to scan; and query fan-out is upgraded, connecting to a broader range of links. Google says Hydrogen is now the default model for the hardest questions in AI Mode and is used for the hardest questions in AI Overviews. In other words, the next time you search on Google, Hydrogen is probably already doing the work behind the scenes, and you will notice nothing — which is exactly how a "daily driver" should behave.
The changes in the Gemini app are closer to daily life: the model follows your guidance more closely, with less back-and-forth; it calls the right tools and skills more often; and it returns things that are ready to use on the first try, in your tone and format and following your specific rules. Google lists the scenarios: drafting email, study guides and rewrites in your voice; pulling context from your uploads, Gmail and Google Drive, plus connected apps like Airtable and Dropbox, and breaking a goal down step by step; everyday prototyping and iteration, such as turning real-time lightning data into an interactive 3D globe. These do not sound like a "model launch" so much as "the assistant got better". That follows from Hydrogen's positioning: it is not built to win papers, it is built to do daily work.
Safety: the faster the model, the worse a fast mistake
The blog devotes a full section to safety, for an obvious reason: the faster an agent model is, the more damage a mistake can do. Google's numbers all cluster around "overreach". Compared with 3.8 Flash, Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, and cuts misaligned actions and task shortcutting by 12x. Faced with impossible or blocked tasks, it is up to 2x less likely to bypass permission boundaries, and it achieves a 96.7% honesty rate on Google's internal agent honesty evaluation. Chain-of-thought legibility is preserved at 99.9%, with zero deceptive reasoning edits across more than 200,000 audited turns. Against indirect prompt injection attacks, Hydrogen is more than 2.3x more resilient on the Gray Swan IPI benchmark. On health questions, false refusals stay under 0.4% on HealthBench Pro, because the safeguards distinguish real biosecurity risks from legitimate medical and scientific questions more precisely.
These numbers need careful reading. All of them come from Google's own internal evaluations, whose methods and samples are not fully public; a claim like "81% less" is measured against 3.8 Flash, not against an absolute safety level. The real risk of a fast model was never in the benchmarks but in the long tail of the real world: a model that calls tools, edits code and sends requests on its own can mangle a repository in minutes once it misunderstands the goal. Google's emphasis on "honesty rate" and "chain-of-thought legibility" shows it knows this: you can only let a model work alone if you can see what it is thinking. For vibe coders, the practical meaning of these safety numbers — if they hold — is that you can let Hydrogen run tests and verify changes inside an isolated workspace with more confidence, instead of watching every step.
Drawing the line against the Gemini stories already published
This story is not the same as the Gemini articles already on this site, and the line is worth drawing. The November 10 story about Gemini 4 Carbon going generally available was about the flagship entering Copilot and its price stepping up from $2 to $4. The earlier Gemini 4 Argon launch story was about the long-reasoning flagship and the robotics direction. The November 11 Gemini Decisions API story was about a model that only judges — answering yes/no and multiple-choice questions without generating text. Hydrogen is the fourth character: the fast everyday execution model. It shares a generation with Carbon and Argon as a division of labour, not a replacement; and it differs from Decision 1 because Decision 1 only answers judgement questions, while Hydrogen is a general model that writes code, runs tools and does research. The four stories together are the complete picture of the Gemini 4 generation.
On the fate of 3.8 Flash, the blog gives no retirement date. Google only calls Hydrogen a major standalone upgrade for teams migrating from 3.8 Flash. According to third-party write-ups of Google's AI Studio pricing page, 3.8 Flash currently costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, and from January 1, 2027 it rises to the same $1.50 and $7.50 as Hydrogen. In other words, the question for 3.8 Flash users is not whether to switch but when — after New Year's Day there is no price advantage left in staying on the old model. That price convergence is not in the blog itself; it comes from pricing-page write-ups, so treat it as a prompt to check the pricing page yourself rather than as Google's own words.
A judgement for vibe coders: when to use Hydrogen
That leaves a direct judgement. Hydrogen fits three kinds of work. First, high-frequency execution inside agent loops: reading files, editing code, running tests, verifying results. These calls are numerous and individually cheap, and running them on a flagship is waste. Second, fast-iteration prototyping: web layouts, interactive demos, data visualisations, where Hydrogen's pitch is more complete results in fewer prompts. Third, the worker role in a multi-model workflow: Carbon sets direction, Hydrogen does the work and verifies it, and Orchestrator handles the handover automatically. The work it does not fit is equally clear: system design that needs deep architectural judgement, stubborn bugs that need long reasoning, and any operation where one mistake is expensive. Those still belong to Carbon and Argon, or at least deserve a human review of Hydrogen's output.
Choosing a thinking level can start simple: default high for everyday changes and batch execution; drop to low or medium for things you can see at a glance, since the tokens and time saved add up; switch to xhigh only for a stuck debugging session or a complex multi-step task, remembering that xhigh costs more, runs slower, and can "overthink" a simple prompt. The worst approach is to use one level for everything — that treats Hydrogen as just another fixed model and wastes its biggest design idea. Google has not published absolute latency or price multiples per level, so this switching strategy currently runs on feel, to be recalibrated once AI Studio measurements appear.
One final judgement. The Hydrogen launch is not earth-shattering in itself: no new model architecture, no parameter count that rewrites the industry. But it marks a shift in where the competition is fought — from "who is smarter" to "whose everyday workflow is cheaper, faster and more reliable". When a vendor starts building a model, a set of thinking levels and an automatic handover system specifically for the scenario of "several hundred calls a day", agent programming has moved from demo to daily routine. And daily routine is the real battleground for vibe coding. Whether Hydrogen holds the default slot will be decided by its rework rate in real repositories, not by launch-day benchmarks. That answer will come from real-world testing in the coming weeks.
What is still unanswered
A few things this announcement does not settle, and they belong here so nobody mistakes them for solved. First, every benchmark figure is Google's own claim; the Artificial Analysis index is published by a third party, but Google is quoting its own reading of it, and no independent replication exists yet. Second, the absolute cost and latency of the xhigh level are unpublished — only that it costs more and runs slower than high, with no numbers attached. Third, Orchestrator's 24% quota saving and 19% speedup were measured on Terminal-Bench 4.0; whether they survive contact with real repositories is unknown. Fourth, Hydrogen inside Antigravity is limited to Pro and Ultra subscribers, so free users cannot use it there yet, and the blog says nothing about Vertex AI support, which enterprise users will need to confirm themselves. Fifth, details such as the model ID gemini-4-hydrogen and a March 2026 knowledge cutoff come from Google's model card and pricing-page write-ups rather than the blog text, so check the official documentation before relying on them. These gaps are not flaws; they are reminders that launch-day information is always incomplete, and good judgement has to wait for measurements.
Comments (0)
Sources
Related articles
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.
On 18 November 2026, Google Research and co-authors posted BuggyMLE to arXiv: bugs injected into 222 working ML repositories, and 14 models slide from 95% on the easiest tier to 9% on the hardest. A 9-billion-parameter model trained easy-to-hard beats Claude Opus 5.5 on the 10-bug tier.