Gemini 4 Hydrogen: Specs, Context Window, Pricing, Benchmarks and Availability

Gemini 4 Hydrogen is a proprietary large language model from Google, generally available since 12 November 2026 under the API identifier gemini-4-hydrogen. Google's Google Cloud documentation describes it as built "for speed and efficiency across everyday coding and agentic workflows", and its model card says it is adapted from Gemini 4 Carbon. This profile collects the published facts on one page: specifications, context and output limits, prices for every Gemini API tier, where the model runs, provider-reported benchmarks, the model card's safety findings, data handling and lifecycle status. It is written for developers, architects, buyers and governance teams who need checked numbers with their sources. Every fact comes from Google's own model pages, developer guide, pricing page, model card and plan pages, listed under Standards and references.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Google's public documentation as of November 2026. Model details change, so confirm against Google's current documentation.

Key takeaways

• Generally available since 12 November 2026. Google Cloud lists gemini-4-hydrogen as GA from that date, and the model card was published the same day; it is the third Gemini 4 model after Argon and Carbon.
• One shared 1,048,576-token window. Input, thinking and output share one limit; a single segment returns up to 65,536 tokens, and long decode extends output towards the full window.
• $1.50 input and $7.50 output per million tokens. That is the Standard paid rate; Batch and Flex are half, Priority is $2.70 and $13.50, and the Gemini API pricing page also lists a Free tier for Hydrogen.
• Knowledge cutoff of March 2026. Both the Google Cloud model page and the model card state it, and the card notes some domains may be less current.
• Below critical capability levels, at alert thresholds. The model card states Hydrogen does not reach the Frontier Safety Framework critical capability levels for CBRN or cyber, but reaches the alert thresholds for both.

What is Gemini 4 Hydrogen?

Gemini 4 Hydrogen is the speed and efficiency tier of Google's Gemini 4 family. The Google AI Studio model page describes it as providing "advanced intelligence with speed and efficiency for everyday coding and agentic workflows". The Google DeepMind model card, published on 12 November 2026, says it is "built for rapid execution across diverse agentic workflows" and positions it alongside the frontier models Gemini 4 Carbon and Gemini 4 Argon, which Google designs for deep, long-horizon reasoning. For architecture, training data, hardware and software, the card refers readers to the Gemini 4 Carbon model card, because Hydrogen is adapted from Carbon.

Hydrogen is a thinking model with four thinking levels. It accepts text, images, audio, video and PDF documents, and returns text. The developer guide describes it as the migration target for gemini-3.8-flash and earlier Gemini 3 Flash models. For the full Google line-up, see Google Gemini Models.

Key facts at a glance

As of 2026-11-15.

Release and version history

As of 2026-11-15.

Google lists a single identifier, gemini-4-hydrogen, with no dated snapshots or variants, and the AI Studio page gives its latest update as November 2026. The Gemini API changelog had no entry for Hydrogen when checked on 15 November 2026; its most recent entry was dated 26 October 2026. Google states no allowlisted cyber variant for Hydrogen on the pages checked; Argon and Carbon each have one.

How much does Gemini 4 Hydrogen cost?

As of 2026-11-15, USD per million tokens, from the Gemini API pricing page. Output prices include thinking tokens. No date-based price change is listed for Hydrogen.

• Audio: no separate audio input rate is listed for Hydrogen; the tier's input rate applies.
• Caching: implicit only, so no cache storage price is listed. Google Cloud states cache writes are billed at 1.25x the standard input price by default.
• Grounding: on paid tiers, Grounding with Google Search and with Google Maps each include 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000.
• Billing scope: the pricing page states that every token the model processes is billed, including thinking tokens.
• Google Cloud: Standard PayGo with Usage Tiers, Priority PayGo, batch inference and Provisioned Throughput; fixed quota is not offered.
• Gemini app: Hydrogen is included on the Free plan ($0), AI Plus ($4.99 a month), AI Pro ($19.99 a month) and AI Ultra (from $99.99 a month, with "highest access").

Worked cost examples

Illustrative arithmetic from the rates above, without caching or grounding charges. Thinking tokens count as output, so real bills vary with thinking level.

For prices across providers see Frontier AI Models Compared.

Context window and output limits

Long decode returns the status values requires_continuation, incomplete and completed. The developer guide advises using it only when a response needs more than 65,536 output tokens, and not for chat, latency-sensitive calls, many short tool turns or prompts near the input limit. See Context Windows Explained.

Modalities and tools

The developer guide's recommended thinking levels:

• XHIGH: deep reasoning, mathematical proofs, scientific analysis and the most complex multi-file engineering.
• HIGH (default): multi-step agentic tool loops, code generation, refactoring and complex multimodal analysis.
• MEDIUM: general coding, tool use, document questions and video understanding where latency matters. Set it explicitly to keep the latency profile of Gemini 3.8 Flash, whose default was MEDIUM.
• LOW: high-throughput classification, routing, summarisation and metadata extraction.

API conventions differ from Gemini 3 Flash models:

• Removed controls: thinking_budget and thinking_level="MINIMAL" return HTTP 400; so do frequency_penalty, presence_penalty and candidate_count above 1.
• Fixed sampling: temperature 1.0, topP 0.95 and topK 64 are fixed and custom values are ignored.
• Thought signatures: pass thought_signature back across all earlier turns. Signatures are bound to the prompt prefix, and the default prefix_mismatch_behavior="ERROR" returns HTTP 400 when earlier turns are edited; set DROP if the harness compacts history.
• Chat history: function responses must match the preceding call's id, name and order, histories cannot end with a model turn, and prefilled model turns are unsupported.

See Reasoning Models Explained and Computer-Use and Browser Agents.

Where is Gemini 4 Hydrogen available?

As of 2026-11-15.

For developer channel security see the Google Gemini API and AI Studio Security Guide and the Gemini Enterprise Security Guide.

Access restrictions

• Standard model: generally available with no allowlist, including on the Gemini API Free tier and the free Gemini app plan.
• Contributor tier: a private, invite-only preview pricing tier in which content is used to improve Google products.
• Cyber variants: Google's model versions page lists allowlisted cyber permissive variants for Argon and Carbon; none is listed for Hydrogen as of 2026-11-15.

Benchmarks (provider-reported, dated)

All scores are reported by Google in the Gemini 4 Hydrogen model card (evaluation results dated November 2026; card published 12 November 2026), with the comparators the card itself uses. Independent leaderboard results are not included because none were checked against their own published pages for this profile. See How to Read LLM Benchmarks.

Comparator scores are as Google reports them in the card.

Model card and system card

Google DeepMind published a Gemini 4 Hydrogen model card on 12 November 2026, linked from the AI Studio model page. The card states that the full Safety and Security Report was "coming soon" when checked. What the card states:

• Intended uses: fast execution alone or as a sub-agent alongside frontier models, iterative repository coding and debugging, open-ended agentic tasks, multimodal knowledge work, web development, writing and research.
• Limitations: hallucinations, ongoing work on jailbreak resistance, occasional slowness or timeouts, higher token use at higher thinking levels, and the March 2026 cutoff.
• Automated safety evaluations against Gemini 3.8 Flash: text-to-text safety +0.80 points, multilingual safety +1.10, image-to-text safety +7.00; tone of refusals -0.65 and unjustified refusals +5.57 are recorded as regressions, particularly for multimodal prompts. The card says these revised evaluations are not directly comparable with earlier cards.
• Frontier Safety Framework (April 2026 version): no material capability change over Gemini 4 Argon, so a full critical capability evaluation was not required. Targeted testing found Hydrogen below the CBRN and cyber critical capability levels but at the alert thresholds for both. Harmful manipulation and ML R&D levels were not reached. Loss of control was not formally tested, and mitigations were applied as if that threshold had been reached.
• Mitigations: pre-training data filtering, post-training alignment, product-level filtering and usage monitoring, with stronger safeguards for CBRN and cyber misuse and ongoing work on indirect prompt injection.

See Model Cards and System Cards Explained and the Frontier AI Model Risk Frameworks Guide.

Data handling

• Gemini API Free tier: the pricing page states that content is used to improve Google's products.
• Gemini API paid tiers: Standard, Batch, Flex and Priority content is not used to improve Google's products.
• Contributor tier: content is used to improve Google's products, under that tier's logs policy.
• Google Cloud controls: data residency, customer-managed encryption keys (CMEK), VPC Service Controls and Access Transparency are listed for online prediction, batch inference and context caching.
• Implicit caching and long decode: caching applies to shared prefixes of 4,096 tokens or more; stored long decode interactions keep server data for 7 days, while stateless loops keep none.

Because the Free tier and paid tiers handle content differently, check which tier each project's API key bills to. See Zero Data Retention for AI and How to set spend limits on LLM API keys.

Deprecation status

Gemini 4 Hydrogen is current and generally available. Google Cloud places it in the table of models with shorter availability periods, with no retirement date announced. That page states that such models remain active until a replacement is released and a retirement date is announced, and that a fixed retirement date gives at least 45 days to migrate. See the AI Model Retirement Guide.

How Gemini 4 Hydrogen differs from other models in its family

As of 2026-11-15. Models listed alphabetically. Sources: Google Cloud model versions page, Gemini API pricing page, Gemini subscriptions page and the Hydrogen model card.

• Positioning: the model card describes Argon and Carbon as frontier models for deep, long-horizon reasoning, and Hydrogen as built for rapid execution, including as a sub-agent working alongside them.
• Lineage: Hydrogen is adapted from Carbon, and the card measures its capability change against Argon.
• Price: Hydrogen's Standard rate is $1.50 and $7.50; Argon's is $2 and $10 until 30 November 2026, then $4 and $20. See the Carbon profile for its rates.
• Thinking: Hydrogen adds an XHIGH level, with HIGH as default.
• Free tier: the pricing page lists a Free tier for Hydrogen; Argon is paid tier only.

See the Gemini 4 Argon and Gemini 4 Carbon profiles, Gemini 4 Carbon vs Argon vs Hydrogen, and Gemini Flash and Flash-Lite Versions Compared for the models Hydrogen replaces in migration guidance.

Practitioner checklist

• Pin gemini-4-hydrogen explicitly and record the adoption date; watch the model versions page for a retirement date.
• When migrating from gemini-3.8-flash, replace thinking_budget with thinking_level and set MEDIUM if you need the earlier latency.
• Remove temperature, top_p, top_k, penalty parameters, candidate_count above 1, explicit caching code and prefilled model turns.
• Pass thought signatures back on every turn, and set prefix_mismatch_behavior to DROP if your harness prunes history.
• Keep regulated or confidential data off the Free tier and the Contributor tier, where content is used to improve Google products.
• Budget with thinking tokens included, and sum usage across long decode segments.
• Confirm the Google Cloud region (global, us or eu) meets data residency needs before production.
• Store Gemini API keys in a secrets manager, restrict them by API and project, and set spend limits.

Standards and references

• Google Cloud: Gemini 4 Hydrogen model page (accessed 2026-11-15)
• Google Cloud: Gemini 4 Hydrogen developer guide (accessed 2026-11-15)
• Google DeepMind: Gemini 4 Hydrogen model card (accessed 2026-11-15)
• Google AI Studio: Gemini 4 Hydrogen model page (accessed 2026-11-15)
• Google AI Studio: Gemini API pricing (accessed 2026-11-15)
• Google Cloud: Model versions and lifecycle (accessed 2026-11-15)
• Google AI Studio: Gemini API changelog (accessed 2026-11-15)
• Google: Gemini subscriptions (accessed 2026-11-15)

Related NHI Mgmt Group resources: Google Gemini Models, Gemini 4 Argon, Gemini 4 Carbon, Frontier AI Models Compared, Google Gemini API and AI Studio Security Guide.