Skip to content
Skip to the story
Dark watercolour astrolabe arcs for the Gemini 4 Hydrogen report

A folkfox special report.

Gemini 4 Hydrogen: 5 essential facts on speed, cost and AI Overviews

Summarise with
Set as preferred source on Google
17 min read

Google has dropped the Flash label, adapted Gemini 4 Hydrogen from flagship Carbon, and wired it straight into global Search today. Here are the verified prices, independent throughput numbers, small-print safety caveats and AI Overview mechanics that actually decide whether to ship it.

By Katie Delaney. Research checked 12 November 2026. The illustrations are conceptual.

The folkfox vixen examining a glowing silver astrolabe beside dark notebooks, moody warm candlelight.

What Gemini 4 Hydrogen changes across code and Search

Every autumn, developer docs leave a clear scent trail before the press notes cool. Gemini 4 Hydrogen ships today across Google AI Studio, the Gemini API, Vertex AI, Google Antigravity, Android Studio, eligible Google Sheets workspaces, the Gemini app and global Google Search.

While external press headlines frame the release as a successor to September's Flash tier, Google DeepMind's own Model Card makes an important architectural distinction: Gemini 4 Hydrogen is adapted directly from flagship Gemini 4 Carbon rather than stretched up from older Flash weights.

At the same time, developer documentation on Vertex AI flags another commercial nuance: gemini-4-hydrogen launches as a short-term generally available release governed by a minimum 45-day retirement notice policy, with a knowledge cutoff locked at March 2026.

  • Unified 1,048,576-token window shares input, thinking and output tokens.
  • Foreground synchronous calls cap at 65,536 output tokens per segment.
  • Background Interactions API and SDK continuation extend decode toward 1M tokens.
  • Native Office and OpenDocument parsing joins PDF, audio and agentic video.
  • Default thinking level shifts to high, rejecting minimal with HTTP 400.

The distribution story also diverges across outlets covering the launch. On Google Search, VP of Product Robby Stein confirmed Hydrogen is rolling out to all users globally, ending the thirty-day window in which Gemini 4 in Search remained gated behind paid AI Pro and Ultra plans.

Inside the Gemini consumer app, however, reporting splits cleanly. Ars Technica describes AI Plus and free accounts as stuck on 3.5 Flash-Lite, whereas 9to5Google details the model picker assigning 3.5 Flash-Lite to quick answers, Hydrogen to everyday help, and routing free users to Hydrogen via Auto selection.

How the Gemini 4 Hydrogen token window shifts migration code

In Google AI Studio's migration notes, upgrading from 3.8 Flash is not a blind string swap. Older Flash separated a 1M input limit from a 64K output budget, whereas Hydrogen shares a single 1,048,576-token ceiling across input prompt, hidden reasoning and final answer.

Teams migrating existing pipelines also have to strip legacy decoding switches. Custom temperature, top_p and top_k values are silently ignored, while explicit context caching (cachedContents), frequency_penalty, presence_penalty, candidate_count above 1 and thinking_level set to MINIMAL trigger hard errors.

The folkfox vixen weighing brass tokens on a small apothecary scale over warm parchment ledgers.

Gemini 4 Hydrogen pricing and thinking tiers tested

On Google's official Gemini API pricing sheet, Gemini 4 Hydrogen pricing locks into Flash-tier unit economics: $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1M context window, with every thinking token billed at the output rate.

That headline number is only half the ledger. Comparing Gemini 4 Hydrogen vs 3.8 Flash on a real reasoning workload exposes through independent Artificial Analysis v5.0.1 testing why nominal per-token rates mislead engineering budgets.

Despite costing $0.75 input and $3.75 output on its temporary 2026 discount rate, Gemini 3.8 Flash (High) consumed 10.2B input tokens and 126M output tokens across the suite, averaging $7.77 per task for an Intelligence Index score of 31. At twice the nominal token price, Gemini 4 Hydrogen (High) finishes in 6.0 minutes at $1.87 per task while scoring 51.

folkfox chart

Artificial Analysis Intelligence Index v5.0.1 across Gemini tiers

Carbon (High)Index 60Hydrogen (Xhigh)Index 54Hydrogen (High)Index 51Argon (High)Index 50Hydrogen (Med)Index 413.8 Flash (High)Index 31Hydrogen (Low)Index 27
Artificial Analysis Intelligence Index v5.0.1 across Gemini tiers
Model TierScore
Carbon (High)Index 60
Hydrogen (Xhigh)Index 54
Hydrogen (High)Index 51
Argon (High)Index 50
Hydrogen (Med)Index 41
3.8 Flash (High)Index 31
Hydrogen (Low)Index 27
Independent AA Intelligence Index v5.0.1 scores comparing Hydrogen thinking tiers against Argon, Carbon and 3.8 Flash. Higher is better.

Source: Artificial Analysis Intelligence Index v5.0.1 and throughput measurements, checked 12 November 2026

Verified Gemini Developer API rates across the current Gemini 4 and 3.8 lineup
Model & TierInput / 1MOutput / 1MImplicit CacheNotes & Expiry
Gemini 4 Hydrogen (Standard)$1.50$7.50$0.075 (95% off)Flat across 1M; free tier rate-limited
Hydrogen (Batch / Flex)$0.75$3.75$0.0375Asynchronous 50% discount tier
Hydrogen (Priority)$2.70$13.50$0.135Dedicated throughput SLA pricing
Gemini 4 Argon / Carbon (Intro)$2.00$10.00$0.10Intro rate expires 30 Nov; $4/$20 on 1 Dec
Gemini 3.8 Flash (Intro)$0.75$3.75$0.075Doubles to $1.50/$7.50 on 1 Jan 2027

Across the four selectable thinking controls, latency and verbosity pull sharply apart. Artificial Analysis records Hydrogen Low at Index 27 (5.81s time to first token, $0.29 per task), Medium at 41 (12.33s, $0.80), default High at 51 (14.96s, 236 tokens per second, $1.87) and Xhigh climbing to 54 at a steep 41.11s wait and 304M output tokens ($3.16 per task).

folkfox chart

Output speed in tokens per second on Artificial Analysis

Hydrogen (High)236 tok/sHydrogen (Xhigh)198 tok/sHydrogen (Low)186 tok/sHydrogen (Medium)177 tok/s3.8 Flash (High)111 tok/sCarbon (High)76 tok/sArgon (High)64 tok/s
Output speed in tokens per second on Artificial Analysis
Model TierThroughput
Hydrogen (High)236 tok/s
Hydrogen (Xhigh)198 tok/s
Hydrogen (Low)186 tok/s
Hydrogen (Medium)177 tok/s
3.8 Flash (High)111 tok/s
Carbon (High)76 tok/s
Argon (High)64 tok/s
Measured generation speed (tok/s) across Hydrogen tiers, Carbon, Argon and 3.8 Flash. Higher is faster.

Source: Artificial Analysis Intelligence Index v5.0.1 and throughput measurements, checked 12 November 2026

Those speed differences explain Google's new Orchestrator mode inside Google Antigravity, which keeps frontier pricing pressures in check by pairing Carbon planning with Hydrogen worker execution. Google reports the pair uses 24% less quota and finishes 19% faster, though note a clean provenance wrinkle: Koray Kavukcuoglu's announcement attributes those percentages to Terminal-Bench 4.0 while the Antigravity team cites Terminal-Bench 4.1.

Search grounding billing and implicit cache rules

Two hidden cost traps sit in the fine print of Google Cloud's Vertex AI developer guide. First, implicit caching activates automatically only once a shared prefix hits at least 4,096 tokens, and explicit cache TTL controls are delayed until 12 January 2027.

Second, Grounding with Google Search shares a 5,000-request monthly free pool across Gemini 3 and 4 models, then bills $14 per 1,000 individual search queries under a 1,500-request daily cap, plus every internal tool declaration and search result injected into the prompt.

The folkfox vixen peering closely through a brass loupe at fine print on an open specification page.

Gemini 4 Hydrogen benchmarks and the safety caveats in small print

Marketing tables tend to hide in the hedgerow behind methodology footnotes. Reading Google DeepMind's four-page evaluation methodology PDF alongside our competitor comparisons reveals where Gemini 4 Hydrogen benchmarks reflect real leaps and where they reflect curated scaffolding.

Against Gemini 3.8 Flash, Google claims Terminal-Bench 4.1 more than triples from 18.2% to 57.0% cheating-adjusted, overreach drops by up to 81%, misaligned task shortcutting falls 12x, and fabricated claims decline more than 5x alongside a 96.7% internal honesty rate. Yet several standard leaderboards remain absent from the release sheet, including GPQA Diamond, LiveCodeBench, MMMU and SWE-bench Verified.

Google DeepMind Model Card self-reported results vs competitor public scores (Nov 2026)
EvaluationGemini 4 HydrogenClaude Sonnet 5.5GPT-6.1 TerraHarness Note
SWE-sweep13.5%3.1%5.8%Self-run via Antigravity CLI 1.4.8
Terminal-Bench 4.157.0%63.9%Not listedCheating-adjusted, avg of 5 runs
Agents' Last Exam v1.171.6%60.4%69.8%Uses 17 Oct task set (pre-26 Oct patch)
AutomationBench-AA76.1%71.8%60.2%Sourced from AA leaderboard
Vending-Bench 2$14,549$11,165$9,107Sourced from Andon Labs leaderboard
OSWorld-2.1 (82 offline)77.6%83.9%74.2%Max of 3 runs; parallel batch calls
HLE-Diamond58.8%53.3%47.3%Sourced from AA leaderboard

Those footnotes matter when comparing across labs. For Agents' Last Exam v1.1, Google kept the 17 October task release without applying the 26 October update; on OSWorld-2.1, Google self-ran both Hydrogen and Sonnet 5.5 over an 82-task offline subset using parallel batch calls and compaction, returning empty strings rather than aborting when safety filters tripped.

Inside the Gemini 4 Hydrogen Model Card, automated safety evaluations highlight an uncomfortable trade-off for production builders. Multimodal image-to-text safety surged +7.00 percentage points over 3.8 Flash and multilingual safety ticked up +1.10 points, while refusal tone softened marginally (-0.65pp).

Simultaneously, unjustified over-refusals climbed +5.57 percentage points on benign or borderline prompts compared with 3.8 Flash, especially in multimodal turns. Meanwhile, the dedicated Safety & Security Report has not shipped today and is labelled simply as coming soon.

Frontier safety alerts and conservative stealth controls

Under its April 2026 Frontier Safety Framework, Google DeepMind ruled that Gemini 4 Hydrogen did not trigger a Material Capability Change over Gemini 4 Argon, skipping a full fresh capability sweep. Targeted testing in sensitive domains still set off amber lights: both biological CBRN uplift and cybersecurity uplift hit their Critical Capability Level (CCL) alert thresholds, though neither crossed the hard CCL line.

For Loss of Control stealth and situational awareness, DeepMind left the Threshold Capability Level untested and instead applied control, monitoring and security mitigations conservatively as if the threshold had been reached.

The folkfox vixen tracing branching amber trails across an enchanted twilight forest map.

Why Gemini 4 Hydrogen Search shifts matter for AI Overviews

For digital marketers and technical SEOs, the model's arrival inside Google Search carries more immediate urgency than any coding leaderboard. By wiring Gemini 4 Hydrogen directly into global Search and default AI Mode routing, Google has handed query fan-out, passage synthesis and link selection to its newest high-throughput engine.

In TechCrunch's coverage, Google highlights four shifts: parsing messy multi-part intent with typos, issuing simultaneous sub-queries to discover deeper pages, rendering visual comparison tables and interactive widgets at Search speed, and strengthening non-English AI Mode nuance.

Read the announcement like a cunning vixen inspecting an empty burrow, however, and what Google omitted stands out sharply. As PPC Land documented today, the Search blog post never once writes the words AI Overviews, supplies zero fan-out query counts, offers zero referral traffic metrics, and skirts how inline calculators and comparison tables cannibalise publisher clicks.

Even on launch day, live Google Search observations confirm why entity clarity and predictive optimisation beat chasing trailing keyword volume. In a check we ran at folkfox across English US SERPs on launch afternoon, trailing keyword tools registered near-zero history for Hydrogen, yet AI Overviews had already fired on launch queries citing nimble specialist explainers like Ars Technica, 9to5Google and PPC Land while blue links still held month-old Argon and Carbon posts. Meanwhile, broad queries on the bare series name still suffered entity pollution, surfacing 1967 NASA Gemini 4 postage stamps beside AI model cards.

Empirical AI Overviews citation mechanics vs launch-day claims (Google Search Central & Ahrefs 2025-2026)
Signal Under TestEmpirical ObservationSource & SampleOperational Implication
Top-10 SERP overlap37.9% in top-10 blocks; 31.0% past rank 100Ahrefs, Mar 2026 (863K SERPs, 4M URLs)Sub-query fan-out decouples AIO links from rank #1
Word length vs citations0.04 Spearman correlation; 53.4% under 1,000wAhrefs (560K AIOs, 174,048 URLs)Padding word count yields near-zero citation lift
JSON-LD schema additionNo uplift across AIO, AI Mode or ChatGPT (-4.6% AIO)Ahrefs, May 2026 (1,885 treated vs 4,000 ctrl)Schema maintains hygiene but does not buy citations
YouTube non-ranking citations18.2% of unranked AIO links are YouTubeAhrefs, Mar 2026 (863K SERPs)Multimodal transcript and description entities count
Official indexing gatePage must be indexed & snippet-eligible; no special tagsGoogle Search Central AI features docsSitemap 404s, JS delays or noindex kill eligibility silently

That empirical bedrock exposes why stale SEO myths waste editorial cycles. Across 560,346 Ahrefs-tracked AI Overviews, raw word count correlated at 0.04 with citations, and more than half of cited pages ran under 1,000 words. In a separate May 2026 controlled study across 1,885 pages adding schema, Ahrefs measured no citation uplift (AIO visibility actually dipped 4.6%), proving JSON-LD reflects clean sites rather than causing citations.

Yet the top-10 overlap collapse tracked by Ahrefs in March 2026 down to 37.9% aligns directly with Google's query fan-out architecture. We have followed similar zero-click pressures across our reporting on AI Overview accuracy audits, branded query capture, self-ranking listicle penalties and GEO click attrition, and Gemini 4 Hydrogen accelerates that shift by splitting messy multi-clause prompts faster at global scale.

Gemini 4 Hydrogen AI Overviews: passage extraction meets crawl hygiene

Google's own Search Central AI features documentation states plainly that there are no secret AI markup tags or markdown magic files required for AI Overviews or AI Mode. A page simply must be indexed, crawlable in robots.txt, snippet-eligible in static DOM HTML, and structured cleanly enough for sub-query passages to verify a specific factual angle.

In client advisory audits and AI Overview citation tracking, the quiet killers remain mundane: broken news sitemaps, bloated client-side hydration hiding facts behind scripts, ambiguous entity naming where a model cannot tell whether "Gemini 4" refers to API software or historical stamps, and reactive articles written three days after early publishers lock up the first fan-out cluster.

The folkfox vixen perched by a twilight hedgerow holding a lantern over five brass compass markers.

Five checks before deploying Gemini 4 Hydrogen and targeting AI Overviews

Whether your studio is swapping gemini-3.8-flash out of an agent worker pool or racing to win citations inside Hydrogen-powered Search results, disciplined engineering beats hype every night of the week. Before rolling live traffic onto Gemini 4 Hydrogen, run these five concrete checks:

  • Audit every 3.8 Flash call site for deprecated caching, minimal thinking and sampling fields.
  • Test Medium and High against Xhigh using per-task cost, not token price.
  • Inspect multimodal refusal rates before exposing edge workflows to stricter safety filters.
  • Publish explicit model-and-modifier hubs (pricing, thinking, errors) matching live query fan-out.
  • Verify static DOM text, sitemap health and unambiguous entity framing inside opening paragraphs.

For AI engineering teams, start on MEDIUM thinking inside the Interactions API and validate structured JSON outputs with native function declarations before permitting escalation to HIGH or Orchestrator pairing with Carbon. Always track Search Grounding query fan-outs separately so a single user turn generating four queries does not silently quadruple your $14-per-thousand tool bill.

For SEO and content teams wanting repeatable inclusion inside Gemini 4 Hydrogen Search experiences, our client advisory team at folkfox pairs predictive topic modelling with technical precision. Instead of waiting weeks for historical keyword dashboards to recognise a product launch, claim concrete sub-intent territory immediately: exact rate cards, SDK error fixes, independent throughput comparisons, and honest safety caveats grounded in verified primary sources, just as our architecture teardowns aim to do without filler.

Why honest small print wins both builders and search engines

Hydrogen's leap from Flash into a fast Carbon-descended workhorse is genuine, particularly when independent Artificial Analysis runs measure a jump from 31 to 51 at one quarter the completed task cost of 3.8 Flash. Still, a short-term 45-day lifecycle notice, missing coding leaderboards, +5.57pp over-refusal behaviour and an unpublished safety report leave real engineering track marks worth inspecting in full daylight.

In both software selection and generative Search visibility, the edge belongs to teams who measure what a model returns on real workloads and organise every answer cleanly enough for a busy human or a fast Search fan-out engine to cite without guessing.

Quick answers

Direct answers on Gemini 4 Hydrogen pricing, benchmarks, migration rules and AI Overviews mechanics, each unpacked above.

How much does Gemini 4 Hydrogen cost per million tokens on the Gemini API?

On the Gemini Developer API paid tier, the model costs $1.50 per million input tokens, $7.50 per million output tokens (including thinking tokens), and $0.075 per million implicit cached input tokens. Batch and Flex tiers halve token rates to $0.75 input and $3.75 output, while Priority costs $2.70 input and $13.50 output. Grounding with Google Search shares 5,000 free requests per month, then bills $14 per 1,000 search queries performed.

Does Gemini 4 Hydrogen replace Gemini 3.8 Flash across Google Search AI Overviews?

Google's Search announcement states that the release is available globally in Google Search for all users, with deeper intent parsing and query fan-out. 9to5Google adds that Hydrogen is the default model in AI Mode and handles harder questions in AI Overviews, though PPC Land notes Google's post never mentions AI Overviews by name or confirms whether every summary switches immediately.

Why can explicit context caching and temperature parameters fail on Gemini 4 Hydrogen?

Hydrogen drops support for explicit context caching (cachedContents and caches.create), relying exclusively on automatic implicit caching from 4,096 shared prefix tokens at a 95% discount. Custom temperature, top_p and top_k values are ignored, while passing frequency_penalty, presence_penalty, candidate_count above 1 or thinking_level set to MINIMAL triggers an HTTP 400 error.

How does Gemini 4 Hydrogen thinking level choice affect speed, cost and accuracy?

Independent Artificial Analysis v5.0.1 runs show Hydrogen scoring 27 at Low ($0.29 per task, 1.3 minutes), 41 at Medium ($0.80, 3.7 minutes), 51 at default High ($1.87, 6.0 minutes at 236 tokens per second) and 54 at Xhigh ($3.16, 13.0 minutes across 304M verbose tokens).

What did Google DeepMind find in Gemini 4 Hydrogen safety and refusal evaluations?

Compared with Gemini 3.8 Flash, automated safety checks recorded +7.00 percentage points on image-to-text safety alongside +5.57 percentage points more unjustified over-refusals and a 0.65-point drop in non-preachy refusal tone. Under the Frontier Safety Framework, Hydrogen triggered biological CBRN and cybersecurity CCL alert thresholds without breaching critical capability levels, and inherited Stealth TCL mitigations from Argon.

What changes for Google Search AI Overviews citations after the Gemini 4 Hydrogen rollout?

Google says Hydrogen expands query fan-out across subtopics and adds more visual tables and in-answer tools, raising zero-click pressure on generic comparisons. Empirical Ahrefs data shows only 37.9% of AI Overview citations sit in top-10 SERP blocks, meaning publishers earn citations by indexing cleanly on neglected long-tail subtopic queries rather than stuffing schema markup.

Work with folkfox

Turn model shifts into measurable Search citations

folkfox helps technical teams pair verified reporting with passage-level SEO and GEO architecture built for AI Overviews.

Sources

Primary documentation, benchmark runs and Search studies for Gemini 4 Hydrogen, verified 12 November 2026.

  1. Google Blog: Introducing Gemini 4 Hydrogen (Koray Kavukcuoglu)
  2. Google Blog: Google Search gets a major upgrade with Gemini 4 Hydrogen (Robby Stein)
  3. Google Developers Blog: Start building with Gemini 4 Hydrogen
  4. Google AI Studio: What's new in Gemini 4 Hydrogen
  5. Google AI Studio: Gemini Developer API pricing
  6. Google Cloud: Developer's guide to Gemini 4 Hydrogen
  7. Google DeepMind: Gemini 4 Hydrogen Model Card (12 November 2026)
  8. Google DeepMind: Gemini 4 Hydrogen Model evaluation methodology (PDF)
  9. Google Antigravity Blog: Gemini 4 Hydrogen and Orchestrator
  10. Artificial Analysis: Gemini 4 Hydrogen (High) analysis
  11. Artificial Analysis: Gemini 4 Hydrogen (Xhigh) analysis
  12. PPC Land: Google Search gains Gemini 4 Hydrogen for all users
  13. Ars Technica: Google rolls out Gemini 4 Hydrogen, finally replacing 3.8 Flash
  14. 9to5Google: Google announces Gemini 4 Hydrogen with significant leaps
  15. TechCrunch: Google's latest Search upgrade promises better answers
  16. Google Search Central: AI features and your website
  17. Google Search Central: Optimizing for generative AI features on Google Search
  18. Ahrefs: 38% of AI Overview citations pull from the top 10 (March 2026)
  19. Ahrefs: We tracked 1,885 pages adding schema (May 2026)
  20. Ahrefs: Short vs. long content in AI Overviews

Reporting: the folkfox newsroom, from the 20 primary sources listed above, checked 12 November 2026. Illustrations are conceptual watercolour plates of the folkfox vixen made for this report. Nothing here is financial or legal advice.

Leave a Reply

Your email address will not be published. Required fields are marked *

The den

Where to?

Pricing

Choose a section. Enter opens it, Escape continues reading.

Cookie preferences

folkfox uses data the way we use strategy: only when it earns its place.