Google Ships Gemini 4 Hydrogen as Its Default Execution Model, Priced at $1.50/$7.50 per Million Tokens (November 12, 2026)

Event overview ​

On November 12, 2026, Google released Gemini 4 Hydrogen. The official blog post was attributed to Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, and a model card was published by Google DeepMind the same day.

Google's positioning is explicit: Hydrogen does not replace the frontier models Gemini 4 Carbon and Gemini 4 Argon, which excel at deep, long-horizon reasoning. It fills the "fast execution" layer — everyday coding, running tests, verifying changes, retrieving material, and chaining tool calls in agentic workflows.

On the same day, Google deployed Gemini 4 Hydrogen to Google Search, replacing Gemini 3.8 Flash as the default model for AI Mode and taking on the most complex AI Overviews questions. Robby Stein, Google's vice president of search, described it on social media as "the biggest model upgrade in Google Search to date" — a statement reported by media, not a line from the blog post.

Key facts ​

Benchmarks as published by Google (Gemini 4 Hydrogen / Claude Sonnet 5.5 / GPT-6.1 Terra):

1. Pushing Gemini 4 down to the execution layer, not stacking another reasoner ​

What matters most here is not the scores but the product layering.

Within the Gemini 4 family, Carbon and Argon target hard problems — long, deep reasoning such as architectural judgement or untangling subtle bugs. Hydrogen targets the labor: writing code, running tests, verifying changes, batch retrieval. Google states the collaboration plainly: a team needs a frontier model for architectural calls and a fast, disciplined execution model to land the work.

Pricing follows the same layering. At $1.50/$7.50 per million tokens across the full 1M context, long context no longer carries a premium. That price sits where high-frequency calls can bear it: execution-layer volume far exceeds reasoning-layer volume, so the unit price has to come down or the agentic cost model does not close.

2. How to read the benchmark table: ahead on coding and knowledge work, behind on terminal and computer use ​

Split the ten benchmarks by win and loss and the picture is more complicated than "leads across the board."

Hydrogen leads clearly on three: SWE-sweep (13.5% against 3.1% and 5.8%), Agents' Last Exam v1.1 (71.6%), and HLE-Diamond (58.8%). All three relate to locating and repairing problems inside a repository and completing end-to-end professional work — exactly the execution positioning.

Hydrogen trails on three: Terminal-Bench 4.1 (57.0% against Claude Sonnet 5.5's 63.9%), OSWorld-2.1 (77.6% against 83.9%), and Terminal-Bench-Science 0.1 (51.4% against 59.9%). Terminal operation and graphical interface use remain stronger on the competing model.

The takeaway: model comparison has moved from a total score to a division of labor. A model can lead decisively on repository coding and trail on terminal interaction at the same time; which conclusion matters depends on what you use it for.

Note the measurement basis. This table comes from Google, and the harness, prompts and scoring are set by the publisher. Implementations of the same benchmark name differ across vendors, so results cannot be summed or averaged across tables. Vending-Bench 2 reports an account balance in dollars, and OSWorld-2.1 is marked as an offline subset; those labels are themselves caveats.

3. Orchestrator: multi-model composition becomes a product feature ​

Alongside Hydrogen, Google introduced Orchestrator in Google Antigravity. It no longer picks one model to run from start to finish; it switches dynamically between Carbon and Hydrogen within a single session.

Google's measurement: on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while consuming 24% less quota on average and finishing 19% faster. For parallel agent workflows, Carbon can coordinate while Hydrogen powers worker and verifier agents, building and testing across isolated workspaces.

The significance is that model composition becomes the default experience. Users no longer decide which model to use at each step; the system switches by task type. If this holds, model value shifts from single-model scores toward cost per completed task after composition.

4. The search impact: a default-model swap matters more than a release ​

On the same day, Gemini 4 Hydrogen succeeded Gemini 3.8 Flash as the default model in Google Search and AI Mode. That move has a far wider blast radius than a model launch.

Search is the distribution entry point. When the default model changes, the wording of AI summaries, the handling of follow-up questions, and the logic behind which sources get cited all shift. For anyone producing content or doing generative engine optimization, the observation baseline has to be rebuilt: which page types the new model prefers to cite, whether structured content is favored, and how multi-turn questions are handled.

Google highlighted three search-side improvements: better intent understanding for complex multi-part questions, connections to a wider range of relevant websites and links, and more visual output — structured answers, comparison tables and custom tools. All three point the same way: results are becoming usable answer components rather than a list of links.

Impact and limits ​

First, competition is shifting from "strongest single model" to "layering plus orchestration." A frontier model for hard problems, an execution model for throughput, and an orchestrator to schedule between them. If validated, this turns model competition into systems-engineering competition rather than a leaderboard race.

Second, the pricing structure deserves separate attention. Flat pricing across a 1M-token context converts long context from a scarce capability into a default. That raises context consumption in agentic applications while pushing down cost per task.

Third, for content producers the immediate implication is volatility: a default-model swap will shift citation behavior in AI search over the short term, so GEO observation baselines need updating rather than carrying over from the previous model.

Limits to state explicitly:

• All benchmark figures and the comparisons — 2.5x faster, fabricated claims down more than 5x, 24% less quota and 19% faster — come from Google's blog and model card and are publisher-run measurements, not independently replicated by third parties;
• The comparison targets are Claude Sonnet 5.5 and GPT-6.1 Terra, with harness and prompts set by Google; same-named benchmarks across vendors are not equivalent;
• Partner testimonials (Box, Pixelcut, STAN, LatchBio, Astrocade, Cartwheel, Discovery Loop, Zingroll) are customer accounts describing individual experiences, not general performance conclusions;
• "The biggest model upgrade in Google Search to date" is a social-media statement by Google's VP of search, reported by media, not the official blog text;
• Google did not disclose training data scale, parameter count, training compute or distillation details;
• This article does not judge the model against competitors overall; it reports the published per-benchmark outcomes.

Official sources ​

• Google blog: Introducing Gemini 4 Hydrogen (November 12, 2026; Koray Kavukcuoglu)
• Google blog: Gemini 4 Hydrogen: Fast agentic development for coders (November 12, 2026)
• Google blog: Google Search gets a major upgrade with Gemini 4 Hydrogen (November 12, 2026)
• Google DeepMind model card: Gemini 4 Hydrogen (published November 12, 2026)
• Related coverage on this site: Gemini 4 Argon launch, Gemini 4 Argon specs and pricing data, Gemini 4 Carbon vs Claude Opus 5.5 vs GPT-6 Astra