Gemini 4 Hydrogen: $1.50/$7.50 per 1M tokens, 1M context, replaces Gemini 3.8 Flash

Google released Gemini 4 Hydrogen on 12 November 2026, a model it calls its "daily driver for agentic workflows". It costs $1.50 per million input tokens and $7.50 per million output tokens, has a 1,048,576-token context window and replaces Gemini 3.8 Flash in the places where Google picks a default model for you. For developers the headline is the price-to-capability step over 3.8 Flash and a list of API behaviour changes that can break a straight model-ID swap.

Key points

• Gemini 4 Hydrogen launched on 12 November 2026 in the Gemini API, Google AI Studio, Android Studio, the Gemini app, Google Search and Google Antigravity; the model ID is gemini-4-hydrogen.
• Paid-tier pricing is $1.50 per 1M input tokens and $7.50 per 1M output tokens (thinking tokens included), with implicit-cache hits at $0.075 per 1M.
• Google says it scores 57% on Terminal-Bench 4.1 against 18.2% for Gemini 3.8 Flash, and 51 on the Artificial Analysis Intelligence Index at its default setting (54 at the new xhigh level) against 50 for Gemini 4 Argon.
• The default thinking level moves from medium to high, and temperature, top_p and top_k are ignored, so upgrading is not only a model-ID change.
• Nearly all performance figures come from Google or its partners; we have not seen independent reproductions yet.

What to do

If you call gemini-3.8-flash today, Google's own migration notes for the Gemini API list these steps:

1. Change the model ID to gemini-4-hydrogen.
2. Remove hard-coded low max_output_tokens caps such as 8,192, or raise them to 65,536 (the default per synchronous call).
3. Re-test your evaluations at medium and high: the default thinking level is now high, which changes latency and token use. The value minimal is rejected with an error.
4. Delete temperature, top_p and top_k; the server ignores them.
5. Remove cached_content fields: explicit caching (cachedContents) is not supported. Implicit caching applies by default to prompts of at least 4,096 tokens.
6. Declare tools natively or use structured outputs instead of prompting for raw-text JSON tool calls, use previous_interaction_id for multi-turn sessions, and preserve thought signatures if you manage history yourself.

Because the cost per task depends on how many thinking tokens a model spends, measure whole tasks rather than token prices; our guide on comparing LLM prices by cost per task shows how.

What happened

Google announced Hydrogen on its blog on 12 November, with a developer post and a Search post the same day. It describes the model as built for rapid execution across agentic workflows, as opposed to Gemini 4 Carbon, its frontier model for deep, long-horizon reasoning. In the Gemini app's model picker, 4 Argon handles code and analysis, 4 Carbon the most complex work, and Hydrogen is labelled "Everyday help", according to 9to5Google's report.

Availability is broad from day one. Developers get it in the Gemini API and AI Studio (including a rate-limited free tier) and in Android Studio. Paid Google AI Pro and Ultra subscribers get it in Google Antigravity. Businesses get it through Gemini for business and in Gemini in Google Sheets for eligible subscriptions. Consumers can pick it in the Gemini app, free users reach it through "Auto" model selection when it fits the prompt, and in Google Search it is the default model for AI Mode and is used for the hardest AI Overviews questions, according to 9to5Google.

Google also dropped the "Flash" brand: Hydrogen takes the slot that Gemini 3.8 Flash held. Google's announcement does not give a deprecation date for 3.8 Flash.

Background

Hydrogen joins Gemini 4 Carbon and Gemini 4 Argon in Google's Gemini 4 line. We covered Gemini 4 Carbon opening to developers at an introductory price, and Google's cyber-focused Carbon variant. Before that, Gemini 3.8 Flash was Google's fast, low-cost tier, and Hydrogen now takes its place.

The numbers

Google compares Hydrogen mainly with its predecessor. On Terminal-Bench 4.1, a benchmark of agentic command-line tasks, it reports 57% for Hydrogen against 18.2% for Gemini 3.8 Flash, more than triple. It also says Hydrogen finishes tasks 2.5 times faster than 3.8 Flash at its default high setting and cuts fabricated claims by more than five times. Partners who tested it reported 2x to 2.8x faster runs and 12% to 23% lower token use on complex graph, SQL and data-heavy workloads.

Gemini 4 Hydrogen57

Gemini 3.8 Flash18.2

Terminal-Bench 4.1 score (Google's figures) (%)

	%
Gemini 4 Hydrogen	57
Gemini 3.8 Flash	18.2

Source: Google, Introducing Gemini 4 Hydrogen

On the third-party Artificial Analysis Intelligence Index the gap to the bigger siblings is small: 51 at the default setting, against 50 for Gemini 4 Argon, rising to 54 at xhigh. Google lists further results: 58.8% on HLE-Diamond (which it says leads its class), 52.3% on Vals Tax Agent Bench (first overall), 55.8% on Vals Legal Research Bench (tied for first), 53.4% on its Multimodal Knowledge Agent benchmark, 71.6% on Agents' Last Exam v1.1, and fifth place on the arena.ai Agent Arena.

Argon50

Hydrogen high51

Hydrogen xhigh54

Artificial Analysis Intelligence Index (as reported by Google) (index points)

	index points
Argon	50
Hydrogen high	51
Hydrogen xhigh	54

Source: Google, Introducing Gemini 4 Hydrogen

The limits matter as much as the scores. The window is one shared 1,048,576-token budget for input, thinking and output. A synchronous call returns up to 65,536 output tokens; longer outputs need background: true with a higher max_output_tokens, up to the window minus your input, or client-side continuation. Carbon and Argon support 32,768 on REST, so Hydrogen allows twice as much per call. Hydrogen defaults to high and supports xhigh, while Carbon and Argon default to medium and stop at high.

How the models split the work

In Antigravity, Google's agentic coding environment, a new Orchestrator can switch between Carbon and Hydrogen inside one session. Google reports "near-Carbon task success rates" on Terminal-Bench 4.0 with 24% less quota consumed on average and 19% faster completion. That is a vendor claim about its own product, and Google does not publish a head-to-head Hydrogen versus Carbon table.

Google also lists safety and reliability claims aimed at agent builders: overreach in agentic tasks reduced by up to 81%, under-service by up to 79%, and up to twice as unlikely to bypass permission boundaries when a task is impossible or blocked. It reports a 96.7% score on its internal agent honesty evaluation and says the model is more than 2.3 times more resilient on the Gray Swan indirect prompt injection benchmark. These come from Google's own evaluations.

What it means

For teams running agents on 3.8 Flash, Hydrogen is a likely upgrade: more capability for a price Google states plainly and a cheaper cache tier. The catches are practical. A default of high thinking can raise per-task cost and latency even when the per-token price looks friendly, so the sensible test is end-to-end cost and success rate on your own tasks at medium and high, not the headline price.

For Google, Hydrogen puts a strong model into every free surface at once: Search, the Gemini app's Auto mode and the free API tier. That raises the baseline users meet by default. For competitors it narrows the gap between a "fast" model and a "frontier" one: the Artificial Analysis gap between Hydrogen and Argon is a single point at default settings.

Developers who relied on sampling controls or explicit caching will have to redesign: those features are gone from this model, which pushes teams towards prompt structure and implicit caching.

What we don't know yet

• Whether the benchmark figures hold up under independent testing: nearly all are Google's or its partners'.
• How Hydrogen compares directly with Carbon, since Google's comparisons are with 3.8 Flash and Argon.
• Whether the release is in preview or generally available: the documentation page we read does not say.
• When Gemini 3.8 Flash will be deprecated, if at all: no date has been published.
• The model's knowledge cutoff, which the documentation does not state.
• Batch pricing, which the pricing page we read does not mention.

What to watch next

Independent results on Terminal-Bench and the Artificial Analysis index at high and xhigh; real cost-per-task comparisons against 3.8 Flash; any announcement of a 3.8 Flash retirement date; and whether Orchestrator-style routing between a frontier and a fast model becomes the default pattern in other coding tools.

Sources

• Google: Introducing Gemini 4 Hydrogen (primary)
• Google: Gemini 4 Hydrogen: everyday coding and agentic workflows for developers
• Google AI Studio docs: What's new in Gemini 4 Hydrogen
• 9to5Google: Gemini 4 Hydrogen launch
• Google: Google Search gets a major upgrade with Gemini 4 Hydrogen