Google launches Gemini 4 Hydrogen for everyday coding and AI agents

What Google launched on November 12

Google announced Gemini 4 Hydrogen on November 12, 2026, making gemini-4-hydrogen available in Google AI Studio and the Gemini API on both free and paid tiers. Standard paid pricing is $1.50 per million input tokens and $7.50 per million output tokens across the full context window. Google describes it as a model for everyday coding, multimodal work and rapid execution in AI-agent workflows, rather than a replacement for every difficult reasoning task.

The developer guide confirms both GenerateContent and the Interactions API. Google also announced a global rollout across Gemini for business, the Gemini app and Search, with Google Antigravity access for paid Google AI Pro and Ultra subscribers. These are product-specific access paths, not one unlimited entitlement: the API free tier is rate-limited, and Google’s app footnote says free users receive Hydrogen through automatic model selection when appropriate for the prompt.

The important buying decision is whether a dedicated execution model can do routine work well enough to reserve more expensive reasoning for difficult steps. Hydrogen can run alone or alongside Gemini 4 Carbon and Argon. Availability and claimed capability justify an evaluation; they do not establish that switching an existing agent will improve its completed-task cost.

A million-token window is one shared budget

The API guide specifies a unified 1,048,576-token window shared by input, thinking and output. That is not a million tokens of input plus another million tokens of output. Filling the window with documents leaves less space for reasoning and the answer.

Synchronous responses support up to 65,536 tokens per hop, whether streaming or non-streaming. To generate more, Google documents background execution with max_output_tokens above 65,536, or client-side continuation in current SDKs. The maximum remains bounded by the unified window minus input tokens. Setting background to true without increasing max_output_tokens still stops at 65,536.

The output limit also includes thinking tokens. An inherited low output cap can therefore truncate a response before the final answer is complete. Review caps together with reasoning settings rather than treating the advertised long decode as the default behavior of a single request.

Four thinking levels change the cost comparison

Hydrogen supports low, medium, high and xhigh, with high as the default. Google recommends reserving xhigh for difficult multi-step problems where high falls short, warning that it increases latency and token use and can overthink simpler prompts. The guide recommends evaluating all four levels and choosing the lowest one that meets the quality requirement.

Standard input costs $1.50 per million tokens, implicit cached input costs $0.075 per million tokens, and output including thinking costs $7.50 per million tokens. Google says implicit caching applies by default to eligible prompts of at least 4,096 tokens, with the discount applying on cache hits. It does not support explicit cachedContents or caches.create. A cache-hit discount is not a promise that every repeated prompt will hit.

Google also says billing counts internal grounding system instructions and function declarations, and that Google Search grounding remains billed per query. Our analysis: compare total billed usage per accepted task, not just the headline input price. A higher reasoning setting, additional grounding queries or failed retries can change the result even when the token tariff is flat across context lengths.

What the performance claims actually establish

Google reports that Hydrogen at its default high setting finishes tasks 2.5 times faster than Gemini 3.8 Flash and scores 57% versus 18.2% on Terminal-Bench 4.1. It also reports more than a fivefold reduction in fabricated claims. These are vendor-reported results. AccessAllGPT did not reproduce them, and they are not a guarantee for a particular repository, tool harness or document workload.

The announcement reports an Artificial Analysis Intelligence Index score of 51 at high and 54 at xhigh, compared with 50 for Argon. We read those figures in Google’s announcement, not an independently retrieved current leaderboard. The different thinking settings matter: the highest score should not be compared with the default setting’s latency as if both describe one configuration.

Google’s new Orchestrator in Antigravity dynamically switches models within a session. Google reports near-Carbon task success on Terminal-Bench 4.0 while consuming 24% less quota and finishing 19% faster on average. That is a separate orchestration result on a different benchmark version, not Hydrogen’s standalone Terminal-Bench 4.1 result or a measured API-bill saving.

What changes for an existing Flash agent

Google recommends Hydrogen as an upgrade from Gemini 3.8 Flash and earlier Gemini models, but its migration checklist goes beyond replacing the model ID. Hydrogen defaults to high rather than medium and rejects minimal. Explicit cached_content fields need removal; temperature, top_p and top_k custom values are ignored by the server.

The guide warns against asking for raw-text JSON tool calls through system instructions with no declared tools, which can produce MALFORMED_FUNCTION_CALL. Declare functions natively or use structured output schemas. For multi-turn sessions, preserve previous_interaction_id in the Interactions API, or prior model turns and thought signatures unchanged in client-managed GenerateContent histories. Do not reuse one thought signature across parallel function-call parts.

For Carbon or Argon users, Google says Hydrogen uses the same request schema, but thinking defaults and supported levels differ. That makes a worker-model trial more straightforward than a Flash migration, not risk-free. Recheck tool behavior and state handling before putting a mixed-model loop behind production permissions.

The next step is a reversible workload trial

Start with one representative public or synthetic coding task and one document or tool-use task from your evaluation set. Compare Hydrogen at medium and high against the existing configuration, holding tool access and acceptance tests constant. Record correctness, interventions, wall-clock time, cache behavior and billed thinking as well as visible output.

If the workflow already uses Carbon, try delegating a bounded edit or verification step to Hydrogen before changing the planner. Keep permission checks, spend limits and rollback outside the model. Google’s documented free tier offers a way to explore access, but we did not call it, establish its account-specific limits or run a hands-on review. Our recommendation is an evaluation path, not a measured endorsement.

Disclosures

AccessAllGPT received no payment, credits or private briefing for this article. Google did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent and is not affiliated with Google. Publication-wide relationships are listed on the disclosures page.