This piece reads only what four Google official pages say word for word, and what they do not say, at two boundaries: an entry-tier handoff and a two-model handoff between Carbon and Hydrogen. As of November 13, 2026, the first-hand sources are the blog.google main launch post, the antigravity.google Orchestrator post, the Google Search upgrade post, and the developer build post, all dated November 12, 2026 verbatim. A next-day timestamp in reposts belongs to the repost, not the release, so the time anchor here stays As of November 13, 2026. There are no rankings, no buying advice, and no integration steps: every official figure is kept exactly as written, with no conversion, no merging, and no conclusion supplied on Google's behalf.
An entry-tier handoff and four thinking levels
Google places Gemini 4 Hydrogen at the entry tier of the Gemini 4 series, taking over the role of Gemini 3.8 Flash. The main launch post does not lean on superlatives; it puts the weight on checkable settings: four thinking levels, verbatim low, medium, high (default), and a new xhigh tier. The parenthetical "high (default)" cannot be dropped, because nearly every speed and score figure below is pinned to a specific setting, and confusing the default tier with the added xhigh tier is the easiest way to misread them. The developer build post calls the same thing four configurable thinking levels, and neither post gives a single cross-tier score. An entry-tier handoff changes the tier, not the series name: Hydrogen stays inside Gemini 4, and the official pages present it neither as a freshly numbered generation nor as a rename of 3.8 Flash.
1M in and out, and the flat pricing wording
The title phrase 1M tokens in and out maps to one sentence in the main launch post: Hydrogen pairs native 1 million-token input with up to 1 million output tokens. Native 1 million-token input describes a native context window, not a figure reached by compression; the output side says up to 1 million output tokens, so 1M on output is a ceiling, and the post does not say input and output are equal or that every run emits a full 1M. Flattening native and up to into "1M input and output both ways" erases a distinction the official sentence keeps, so this piece keeps native input and up to output separate in the body.
Pricing is bound to the window in the same post, verbatim: 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. The load-bearing phrase is flat across its entire 1 million-token context window: the whole 1M-token window is priced the same way, with $1.50 per million input tokens and $7.50 per million output tokens, and is not broken into priced tiers. Those two figures are per-million-token unit prices, not a flat fee per task; "flat across the window" means pricing does not segment by position in the window, not that long and short contexts cost the same. Reading flat as one fixed price would collide with the per million unit that follows. The first-hand source for both prices is the body of the main launch post; the official pricing page was not retrieved, so this piece does not treat it as the source. The developer build post lists no price, token ceiling, multipliers, or benchmark scores, and the two pages cannot fill each other's gaps.
The 57% coding benchmark: Hydrogen standalone, not Orchestrator
The main launch post puts speed and the coding benchmark in one sentence: finishing tasks 2.5 times faster at its high default setting while more than tripling Terminal-Bench 4.1 scores (57% vs. 18.2%), cutting fabricated claims by more than 5x. Three numbers need separate handling. 57% and 18.2% are a Terminal-Bench 4.1 comparison between Hydrogen standalone and Gemini 3.8 Flash, which the post calls more than tripling Terminal-Bench 4.1 scores; the 57% belongs to Hydrogen as a single model, not the two-model system below. "Finishing tasks 2.5 times faster" carries the official qualifier at its high default setting, so generalizing it to "2.5x faster overall" drops the boundary Google drew. Fabricated claims are cut by more than 5x, that is more than fivefold, and must not be downgraded to an approximate multiple; more than recurs as a lower-bound hedge.
Placing 57% vs. 18.2% correctly is the tightest call here. It is a Terminal-Bench 4.1 comparison between two standalone models, Hydrogen at 57% and 3.8 Flash at 18.2%. The Antigravity post writes the same comparison as 57% on Terminal-Bench 4.1 (more than tripiling 3.8 Flash's 18.2%), where tripiling is that page's own spelling, kept as-is rather than corrected. Moving these onto Orchestrator merges two different objects, the slip common in second-hand retellings. What Google gives is one paired score on Terminal-Bench 4.1, not a cross-benchmark ranking or a blanket success rate; this piece draws no ranking from it. The main launch post also lists 2x to 2.8x faster and 12% to 23% fewer tokens as separate official figures, each valid in its own context and not to be blended with 2.5 times faster into one uniform multiplier.
Safety, honesty, and the intelligence index
The main launch post also lists, verbatim: It reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, and cuts misaligned actions and task shortcutting across tool-use and coding evaluations by 12x. The up to cannot be dropped: 81% and 79% are upper bounds, and rewriting them without up to as fixed values turns a hedged statement into a fixed one. The 12x is the official by 12x, a different quantity from the two up to items and not to be merged into one percent. Other verbatim figures include achieves a 96.7% honesty rate, preserves 99.9% chain-of-thought legibility with the qualifier across more than 200,000 audited turns, and more than 2.3x more resilient on the Gray Swan IPI benchmark.
The intelligence index figures need their source named. The main launch post writes 51 vs. 50, and reaching 54 at its highest xhigh thinking setting; the Antigravity post writes scoring 51 on the Artificial Analysis Intelligence Index (surpassing Argon's 50, and reaching 54 at its highest setting). Both refer to the Artificial Analysis Intelligence Index as quoted by Google's official blog, not to an independent test Google ran itself. The numbers are tied to Argon's 50 and to the highest xhigh setting, so both the quote and the tier are load-bearing; calling it an official first place would overreach the wording. Further official scores such as HLE-Diamond (58.8%), Vals Tax Agent Bench (52.3%), and #5 on Agent Arena live on different benchmarks and are not folded into one table, so this piece makes no cross-benchmark comparison.
Orchestrator and the 4.0 / 4.1 labeling split
The Antigravity post describes a two-way handoff between Carbon and Hydrogen: Carbon does read-only exploration, sets a plan, and predefines replanning triggers; Hydrogen writes code, changes dependencies, and runs tests; when blocked, control is handed back to Carbon to revise the plan. Routing is not fixed from the first prompt but adjusts against those triggers during execution. Orchestrator is where this handoff lands, written as near-Carbon task success rates, meaning the two-model collaboration approaches Carbon's task success while saving quota and time. The object is a collaboration system, not a single model's score, and differs from the 57% vs. 18.2% standalone comparison above.
The point that must be written exactly as found is that two official pages label the same numbers with different benchmark versions. The blog.google main launch post reads: 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.google Orchestrator post writes the same 24% less quota, 19% faster, and near-Carbon as: On Terminal-Bench 4.1, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster. So the Google main launch post labels it Terminal-Bench 4.0, while the Antigravity blog labels it Terminal-Bench 4.1: the same numbers, the same wording, two official pages, two version tags. This piece does not pick one, does not reconcile them, does not infer which version Google meant, and does not flatten both into a "4.x"; it sets out the split as found, main launch post 4.0 and Antigravity blog 4.1.
Two confusions stay apart. 24% less quota and 19% faster are verbatim figures from Google's own internal testing of Orchestrator, written in both pages with the on average qualifier, describing quota consumption and completion speed, not the 57% and 18.2% coding scores; attaching them to 57%, or treating 57% as an Orchestrator result, merges the two objects. The version split attaches only to the Orchestrator set of near-Carbon plus 24% and 19%; the standalone 57% vs. 18.2% is labeled Terminal-Bench 4.1 in both pages and has no 4.0 / 4.1 wobble. A second-hand line saying "57% on 4.1, near-Carbon on 4.0" folds the standalone comparison and the collaboration system's version tags into one. This piece keeps them separate: standalone comparison on 4.1, collaboration system's 24% and 19% carrying the 4.0 versus 4.1 split.
Rollout and what each of the four pages says
On rollout, the Antigravity post writes: Gemini 4 Hydrogen is rolling out across paid Google AI Pro and Ultra subscribers, with Orchestrator opening first to Google AI Ultra. Rolling out describes a staged push, paid limits it to paying subscribers, and Orchestrator and Hydrogen do not open to identical audiences; none of this compresses into "available to everyone at once". The Search upgrade post, signed by Robby Stein, VP of Product, Google Search, contains verbatim This is the biggest model upgrade to Search yet., along with Hydrogen is now available globally in Google Search for all users and across all supported AI Mode languages; the first sentence is Google's self-description, quoted as such and not adopted as this piece's absolute judgment. The developer build post writes four configurable thinking levels and lists Gemini API, AI Studio, Android Studio, Gemini for business, and Antigravity (Pro, Ultra), with no price, token ceiling, multipliers, or benchmark scores.
Spreading the four pages out shows which claims have a source: pricing, token ceilings, tier scores, safety figures, and the intelligence index quote sit in the main launch post; the handoff mechanism, 24% and 19%, and rollout sit in the Antigravity post; global Search availability and language coverage sit in the Search upgrade post; configurable levels and entry points sit in the developer build post. Each number stays next to its page. The official pricing page and developers.googleblog.com were not retrieved, so this piece does not present them as published sources; the first-hand basis for the prices is the body of the main launch post.
What the official pages do not mention
Anything the four pages do not mention is written as "not mentioned in the official pages", never as a negative claim made on Google's behalf: the former reports that something was not read in these pages, the latter would put a denial in Google's mouth that Google never made. The Search upgrade post writes available globally in Google Search for all users and across all supported AI Mode languages, and goes no further into a default model choice or a routing arrangement for the hardest questions; this piece does not expand those points or turn "these pages do not say" into an official negative statement. The developer build post lists no price, token ceiling, multipliers, or benchmark scores, which says only that the page omits them, not that Google has no such terms elsewhere.
The same boundary governs the reverse temptation to invent what was not published: specs, tier-by-tier comparisons, and launch timelines without verbatim official support are not filled in, and second-hand retellings never stand in for a page that was not retrieved. Reading four official pages word for word establishes only the numbers and phrasing they put in print and the qualifiers they left in. Marking blank spaces as "not mentioned in the official pages" stays closer to the facts than supplying a tidy conclusion.
A different angle from existing in-site pieces
Several in-site Gemini 4 pieces already exist, and this one takes only the entry-tier and two-model handoff. The Gemini 4 Carbon piece and the pricing-window SOP follow the generative-model launch and its pricing window; the decision-model scoring piece covers a different kind of model; the robotics and AlphaProof hotspot, the model comparison, and the pricing comparison each keep their own angle. The 57% coding benchmark, the 1M tokens in and out, and Orchestrator's 24% and 19% with its version split belong to this entry-tier line and do not overwrite those earlier arguments.
Related reading
- Gemini 3.8 Flash: the previous entry-tier model, useful for following the handoff line.
- API pricing comparison: a cross-vendor look at pricing terms, a companion angle to the verbatim pricing above.