Google Releases Gemini 4 Hydrogen With Dynamic Thinking and New Antigravity Agent Control

Google DeepMind on this Wednesday released Gemini 4 Hydrogen, the third and final model in its Gemini 4 family, completing a lineup that began with Argon in September and Carbon in October. Available immediately to all users across Google's consumer and developer surfaces, Hydrogen is designed for everyday coding, agentic automation, and high-volume API deployments — filling the slot formerly occupied by Gemini Flash while retiring that branding entirely. Google claims Hydrogen exceeds Gemini 4 Argon at the default thinking setting on the Artificial Analysis AI Intelligence Index; independent data from Artificial Analysis places Argon at 53, meaning Hydrogen at its default "high" setting — where it scores 51 — registers below Argon by that measure. Only at Hydrogen's maximum "xhigh" compute tier, which reaches 54, does it exceed Argon's independently measured score. On the separate Terminal-Bench 4.0 agentic benchmark maintained by vals.ai, Hydrogen does not yet have a scored entry as of launch; that leaderboard is led by Anthropic's Claude Fable 5.2 and Claude Opus 5.5, tied at 65.15 percent, ahead of Claude Sonnet 5.5, Kimi K3.1 Pro, and Gemini 4 Carbon.

Hydrogen Completes Gemini 4's Element-Named Tier Structure

The release closes out what Google has constructed as a full-spectrum model family, each named after a chemical element rather than a marketing descriptor. Gemini 4 Argon, launched publicly on September 30, 2026, anchors the frontier reasoning tier. Gemini 4 Carbon was announced October 28 with restricted Fairwind access before reaching broader availability on November 10, extending that frontier capacity into long-horizon agentic work and cybersecurity tasks.

Hydrogen occupies the speed-and-breadth tier — what Gemini 3 Flash once held by name but Gemini 4 now reframes through the lens of chemistry's lightest element: fast, abundant, and foundational. The retirement of "Flash" as a product tier signals that Google is moving away from tiered marketing names (Flash, Pro, Ultra) in favor of a naming convention that emphasizes the model's generation and position within a family without implying speed as a constraint on quality.

That framing reflects a genuine technical shift. Hydrogen is not simply a faster, cheaper version of Argon with capped capabilities. At its default setting — labeled "high" thinking — the model scores 51 on the Artificial Analysis AI Intelligence Index. Artificial Analysis independently measures Argon at 53; Google's launch documentation cites Argon at 50. Using the independent figure, Hydrogen at default falls two points below Argon. Push Hydrogen to its maximum setting, "xhigh," and the index rises to 54, placing it one point above Argon's independently measured score. The result is more nuanced than Google's framing suggests: Hydrogen exceeds Argon on this index only at maximum compute, not at the default setting.

Four Thinking Levels Bring Predictable Cost-Quality Control

The most architecturally significant feature in Hydrogen is a formal system of four discrete thinking levels: low, medium, high, and xhigh. The default is high.

Thinking levels are not new in the broader industry — Anthropic introduced extended thinking in Claude 3.7 Sonnet in early 2025, and reasoning modes have proliferated across frontier labs since. What Gemini 4 Hydrogen formalizes is a discrete four-tier structure rather than a continuous slider or a simple on/off toggle, allowing production systems to select a specific compute budget without manual calibration.

At the high setting, the model applies an extended internal reasoning pass before returning a response — analogous to chain-of-thought inference, but baked into the model's generation process rather than appended as a prefix prompt. At the low setting, reasoning overhead is minimized, prioritizing raw throughput. At xhigh, the model deploys its maximum available reasoning budget, which in Hydrogen's case adds three index points on the Artificial Analysis benchmark at the cost of additional latency.

Google has not disclosed the token budgets associated with each level, nor has it published architecture details for Hydrogen — including whether the model uses a mixture-of-experts design that would allow selective expert activation to manage per-token compute cost. With MoE architectures, only a subset of parameter blocks activate for each inference step, allowing a model to maintain a large total parameter count while keeping per-token compute tractable. If Hydrogen uses this approach, the thinking levels may also interact with how many experts are activated, not just how long the reasoning chain runs. The behavior and pricing of thinking tokens relative to output tokens at each setting has also not been fully specified in the launch documentation.

The absence of explicit caching — where developers manually tag context blocks to be cached and reused at a discounted token rate — is a notable omission at launch. Gemini 3.8 Flash offered explicit caching as a cost-reduction mechanism for applications that repeatedly send the same large system prompt, such as a long coding context or document corpus. Hydrogen ships with automatic caching only. Google may add explicit caching in a subsequent release, but its absence shifts the caching optimization burden: developers cannot currently guarantee that a large context will be cached, only that it will be cached if it exceeds the 4,096-token threshold.

Hydrogen arrives with a 1,048,576-token shared context window — equal to that of Gemini 4 Carbon — encompassing input, output, and thinking tokens together. This shared-window design has practical consequences for extended agentic sessions: a workflow that generates a long reasoning trace at xhigh thinking level consumes context window capacity that would otherwise be available for additional tool outputs or conversation history. Developers building long-horizon agent loops will need to account for thinking-token consumption when estimating context budget, particularly at higher thinking levels where the reasoning trace is longer. Automatic key-value cache activates at 4,096 tokens, down from the 6,144-token threshold used by Gemini 3.8 Flash. For multi-turn agentic sessions that cycle through the same system context repeatedly, that lower threshold reduces effective inference cost modestly, since cached context does not incur full token charges on subsequent turns.

Terminal-Bench Shows Hydrogen in Mid-Field, Not at the Top

Google's own benchmark framing compares Hydrogen against Gemini 3.8 Flash, its direct predecessor. On Terminal-Bench 4.1, a Laude Institute benchmark for AI agents operating in terminal environments, Gemini 3.8 Flash scores 18.2 percent. Hydrogen scores approximately 57 percent — a dramatic improvement that the launch materials emphasize.

What those materials do not feature is where Hydrogen sits in the broader competitive leaderboard.

The separate Terminal-Bench 4.0 leaderboard maintained by vals.ai covers 66 terminal tasks across seven domains: software engineering, scientific computing, machine learning, operations, security, hardware, and media processing. Each task runs five trials with an eight-hour agent timeout. The benchmark is methodologically transparent — task descriptions, scoring code, and results are publicly posted — making it one of the more credible independent measures of agent capability in a software environment. Terminal-Bench 4.0 and Terminal-Bench 4.1 are separate evaluations with different tasks; Google's self-comparison between Hydrogen and Flash uses Terminal-Bench 4.1 scores, while the independent competitive leaderboard operates on Terminal-Bench 4.0.

As of November 9, 2026, the vals.ai Terminal-Bench 4.0 leaderboard places Claude Fable 5.2 and Claude Opus 5.5 tied at the top at 65.15 percent, followed by Claude Sonnet 5.5 at 64.14 percent, Kimi K3.1 Pro at 63.64 percent, and Gemini 4 Carbon at 63.13 percent. Hydrogen does not yet appear on that leaderboard — the model launched after the most recent scoring update — meaning its Google-reported Terminal-Bench 4.1 score of approximately 57 percent cannot be directly compared to these Terminal-Bench 4.0 results. The directional picture nonetheless suggests Hydrogen trails the competitive mid-tier field on agentic terminal tasks; a formally comparable Terminal-Bench 4.0 result for Hydrogen is not yet available.

The implication is not that Hydrogen is weak. A 57 percent success rate on Terminal-Bench 4.1 — a benchmark where Gemini 3.8 Flash scored 18.2 percent — represents substantial generational progress. The implication is that developers choosing a model specifically for autonomous terminal-based software engineering workflows should expect to find faster and more capable alternatives available from Anthropic and from Google's own Carbon model, based on the Terminal-Bench 4.0 leaderboard. For the agentic coding use case Hydrogen targets, pricing rather than benchmark performance is Hydrogen's principal competitive advantage over Sonnet 5.5.

Google also reports Hydrogen's performance on a suite of specialized agentic evaluations: 52.3 percent on Vals Tax Agent (claimed first place), 55.8 percent on Vals Legal Research (tied first), 71.6 percent on Agents' Last Exam v1.1, and 53.4 percent on the MKA benchmark. These are company-reported figures. The Vals and MKA benchmarks target professional-domain task completion — legal research, financial analysis, structured knowledge retrieval — where Hydrogen's claim to leadership has not yet been independently reproduced. Google's launch documentation covers the full benchmark suite.

Antigravity Orchestrator: Dynamic Model Switching for Agentic Sessions

The most architecturally novel feature in the Hydrogen launch is not the model itself but what it enables inside Google's Antigravity developer platform.

Antigravity, Google's agent-first IDE that reached version 2.0 at Google I/O in May 2026, now ships an Orchestrator component that dynamically selects between Gemini 4 Carbon and Gemini 4 Hydrogen within a single agentic session. Rather than a developer committing to one model for an entire workflow, the Orchestrator routes individual tasks or subtasks to whichever model it determines is better suited — routing complex reasoning steps to Carbon and high-volume, faster-turnaround execution steps to Hydrogen. The Hydrogen launch announcement details the Orchestrator architecture and integration requirements.

Google reports that Orchestrator-routed sessions reduce model quota consumption by 24 percent and complete tasks 19 percent faster compared to running the same workflow entirely on Carbon. Both figures are company-reported and have not yet been independently verified. The 19 percent speed improvement likely reflects the Orchestrator's ability to route tasks to Hydrogen when Carbon's additional reasoning overhead is unnecessary — Carbon is slower at inference, and not every agent step requires frontier-tier reasoning. The 24 percent quota reduction likely reflects the lower token pricing on Hydrogen calls that make up a portion of the session. The Antigravity changelog covers Orchestrator implementation details for developers.

Orchestrator is rolling out to Google AI Ultra subscribers and enterprise accounts initially, with broader availability unspecified.

The Orchestrator concept matters beyond its immediate productivity figures. Routing between models dynamically within a session is architecturally distinct from a user manually choosing a model at session start. It implies that the optimal model for a task can change mid-workflow — that a single session might start with Carbon to reason through an architecture decision, switch to Hydrogen for rapid implementation, and return to Carbon for a final quality review — without the developer scripting those transitions explicitly. If the routing logic is accurate, Orchestrator could become the primary interface through which many developers interact with Gemini 4 models, abstracting the choice of which model to use into the platform rather than leaving it as a manual decision.

Pricing: Flat-Rate Structure Masks a Price Increase from Flash

Hydrogen's pricing is straightforward: $1.50 per million input tokens and $7.50 per million output tokens, with the same rate applying throughout the full 1M-token context window. No tiered pricing by context length or thinking level is currently applied.

The input price of $1.50 per million tokens matches Gemini 3.8 Flash exactly. The output price of $7.50 per million tokens is exactly double the Gemini 3.8 Flash output price of $3.75. For workloads with high output-to-input ratios — typical of agentic code generation — the effective cost of running Hydrogen is substantially higher than running the prior-generation Flash model it replaces, particularly before automatic caching reduces the per-turn cost.

The comparison to Argon and Carbon is more favorable. At launch, Argon and Carbon are priced at $2 per million input tokens and $10 per million output tokens. After November 30, 2026, those prices are scheduled to rise to $4 input and $20 output — making Hydrogen comparatively cheaper than the frontier tier by a widening margin. For workloads that do not require Carbon or Argon's depth of reasoning, Hydrogen's $7.50 output rate will represent meaningful savings against a December 2026 Carbon cost of $20 per million output tokens. Full pricing details are available in Google's Hydrogen launch announcement.

Safety Claims Require Independent Verification

Google reports a cluster of safety metrics for Hydrogen that it characterizes as improvements over its previous generation: 81 percent reduction in agent overreach, 96.7 percent honesty rate, 99.9 percent chain-of-thought legibility, and 2.3 times greater resilience to indirect prompt injection attacks. All are company-reported and evaluated using Google's internal testing protocols. The Gemini 4 Hydrogen model card documents the full safety evaluation methodology.

The agent overreach figure is directionally meaningful — overreach, in which an AI agent takes actions beyond the scope of its assigned task, is a documented failure mode in deployed agentic systems, and an 81 percent reduction would be significant if independently reproduced. Prior-generation Gemini models in agentic configurations were noted by developers to occasionally execute unintended file writes, API calls, or loop iterations beyond the scope of an assigned task; a structural improvement here would directly address a friction point in production deployments. Chain-of-thought legibility at 99.9 percent claims that the model's visible reasoning trace is comprehensible and corresponds to its actual reasoning process — a claim that remains difficult to verify externally without access to the internal architecture. The concern is a known one: reasoning traces in extended-thinking models can become decoupled from actual computation, producing plausible-sounding explanations for decisions made through opaque paths.

No third-party red-team report or safety audit has been published alongside the Hydrogen launch. Independent evaluation of these safety properties, particularly prompt injection resilience and agent overreach behavior in live production conditions, should be expected in the weeks following release as researchers and enterprise security teams test the model.

Where Hydrogen Fits in a Saturating Mid-Tier Field

The mid-tier AI model segment is increasingly competitive, and pricing differences between top contenders are narrowing while performance gaps persist. Claude Sonnet 5.5, released by Anthropic on September 28, 2026, is priced at $2 per million input tokens and $10 per million output tokens — more expensive per token than Hydrogen. Sonnet 5.5 scores 64.14 percent on Terminal-Bench 4.0, while Hydrogen's Google-reported Terminal-Bench 4.1 score of approximately 57 percent is on a separate benchmark iteration, precluding a direct side-by-side comparison. For a workflow generating one million output tokens, Sonnet 5.5 costs $10.00 versus Hydrogen's $7.50 — a 25 percent cost savings for Hydrogen that must be weighed against the terminal task success-rate difference evident across both benchmark versions. OpenAI's GPT-6.1 Terra, released October 26, 2026, occupies a comparable position in OpenAI's lineup; it currently ranks twelfth on the vals.ai Terminal-Bench 4.0 leaderboard at 49.49 percent, priced at $1 per million input tokens and $5 per million output tokens.

Kimi K3.1 Pro's fourth-place standing on Terminal-Bench 4.0 at 63.64 percent deserves attention. Kimi, developed by Chinese AI lab Moonshot AI, has not yet achieved the same distribution footprint as Anthropic or Google models in Western enterprise deployments, and its API access patterns differ. But its Terminal-Bench 4.0 standing is an independent benchmark result — not a company-reported metric — and signals that the competitive frontier for agentic coding work is wider than the dominant US lab narrative suggests.

Hydrogen's most defensible competitive position is not benchmark performance but ecosystem integration. It is the only model that ships with Antigravity Orchestrator support at launch. It is immediately available in Google Search, Google Sheets, Android Studio, and AI Studio without additional configuration. For developers already committed to Google's toolchain, Hydrogen's combination of Antigravity integration and the Orchestrator's dynamic routing makes it strategically attractive even where raw benchmark comparisons favor alternatives.

Hydrogen is also the first Gemini 4 model to roll out simultaneously across consumer surfaces, developer APIs, and enterprise deployments. Argon and Carbon launched with more restricted access — initially reaching enterprise and AI Ultra subscribers before expanding. Hydrogen's day-one availability in Google Search and the Gemini app for all users, including free-tier accounts, is itself a deployment decision with strategic implications: Google is positioning its mid-tier Gemini 4 intelligence as the default AI experience for general users, not a premium tier requiring subscription. That represents a meaningful change in how intelligence tiers get distributed across a platform.

The Gemini 4 element naming convention — Argon, Carbon, Hydrogen — also signals that this family is not yet complete in a different sense. The periodic table has more elements, and Google's choice of a naming structure with obvious room for expansion may be the clearest indication of where the Gemini 4 roadmap points: not a final triumvirate of models, but a foundation for further differentiation as agent architectures mature and new capability tiers emerge.

For now, Hydrogen's clearest contribution is not outperforming its competitors but raising the floor: a mid-tier model that scores above a prior-generation frontier model on an independent benchmark at maximum compute, integrates with an intelligent routing layer that reduces overhead in agentic workflows, and ships today across nearly every surface Google operates. Whether Orchestrator's routing efficiency holds in production conditions — and whether the safety claims survive independent scrutiny — will define whether this launch represents a durable advance or a well-framed benchmark moment.