Editorial Team
9 min read
Google launched Gemini 4 Hydrogen and automatic model switching in Antigravity, claiming near-Carbon task success with less quota and faster completion. Lower token rates support the efficiency pitch, but reasoning settings, mixed competitive results and inconsistent benchmark labels leave the cost of completed work to be tested.
Google released Gemini 4 Hydrogen on November 12, 2026, and introduced automatic switching between it and Gemini 4 Carbon in Antigravity, Google's software-development platform. Hydrogen is the latest member of Google's Gemini AI model family, designed to execute coding and other tasks using tools; Carbon supplies deeper reasoning. The company's announcement positions the new Orchestrator option as a way to combine them within one session.
The commercial proposition is to spend expensive reasoning selectively: let Carbon investigate a difficult problem, delegate edits and tests to Hydrogen, then return to Carbon when execution stalls. That makes the relevant buying question the cost and time needed to finish acceptable work, including any handoffs or retries.
Google’s demonstration shows Hydrogen selected for a dependency-update prompt in Antigravity. Source: Google. Source: Gemini 4 Hydrogen and Orchestrator.
Hydrogen is adapted from Carbon and accepts text, images, audio and video, producing text output. Its model card describes rapid execution as its role, while Carbon and Gemini 4 Argon handle deeper, longer-running reasoning. Google’s release history puts Argon’s general availability on October 13 and Carbon’s on November 10. Gemini 3.8 Flash, the earlier model Google recommends developers upgrade from, became generally available on September 2.
Antigravity itself launched in public preview on November 18, 2025, combining an AI-powered development environment with browser control and agents that could plan and execute software tasks. Its Manager view already coordinated multiple agents across workspaces. The launch also offered Google's Gemini 3, Anthropic's Claude Sonnet 4.5 and OpenAI's GPT-OSS models. Multi-agent work and model choice therefore predate Hydrogen; the new feature automates the Carbon–Hydrogen handoff within a task.
Google says static routers assign a model based on the opening prompt, while Orchestrator can move work in both directions as difficulty changes. Its example has Carbon inspect dependencies and plan, Hydrogen make changes and run tests, and Carbon resume if the plan needs repair. Developers can also assign model roles explicitly through the Antigravity SDK, including Hydrogen workers alongside Carbon coordination or review.
Google’s Antigravity announcement says Orchestrator achieved near-Carbon task success while using 24% less quota and finishing 19% faster on average. The benchmark label is inconsistent: that post identifies Terminal-Bench 4.1, while the main model announcement attaches the same figures to Terminal-Bench 4.0. These are company-reported benchmark savings; quota consumption does not establish a customer's dollar saving.
Google’s published token prices; Carbon’s introductory rates apply through November 30, 2026. Hydrogen output pricing includes thinking tokens. These rates do not measure completed-task costs. Source: What's new in Gemini 4 Hydrogen.
Hydrogen’s API documentation lists flat paid pricing across its full context window: $1.50 per million input tokens and $7.50 per million output tokens, including thinking tokens. Both rates are 25% below Carbon’s introductory $2 input and $10 output rates listed in the release notes. That comparison applies to Carbon’s promotion through November 30, 2026, and measures price per token, not cost per completed task.
Tokens are the units used to meter model inputs and generation, including its internal reasoning. Hydrogen offers low, medium, high and a new xhigh thinking level, with high as the default. That changes the starting point for users migrating from Gemini 3.8 Flash, whose default is medium. Carbon and Argon also default to medium and cap at high. Google reports an Intelligence Index score of 51 for Hydrogen at default high, against Argon’s 50; that is a reported comparison of model configurations, not proof of superiority on every task.
Google explicitly cautions that xhigh increases latency and token use and can overthink simple prompts. Its documentation still recommends Gemini 3.5 Flash-Lite for the fastest, lowest-cost classification and routing workloads. Hydrogen's broader capabilities do not remove that cheaper alternative.
The independent model evaluator Artificial Analysis's results at xhigh illustrate the trade-off. It gives Hydrogen an Intelligence Index score of 54, but records 304 million output tokens across the evaluation, compared with a median of 72.8 million among models in a similar price tier. Its weighted cost, accounting for input, caching, reasoning and answer tokens, is $3.16 per index task. Estimated time per task is 13.0 minutes, close to the comparison median of 12.9 minutes.
Those task-time figures are calculated from output-token use and measured output speed. The index combines nine evaluations covering different capabilities. It tests a different scope and reasoning setting from Google's claim that Hydrogen at default high finishes tasks 2.5 times faster than Gemini 3.8 Flash. Google also says early partners saw speeds of 2 to 2.8 times Flash's and token-use reductions of 12% to 23% on complex graph, SQL and data-heavy enterprise workflows. Those company-reported results concern selected workloads, while the independent figures concern xhigh across a broader index. Neither measures the Carbon–Hydrogen Orchestrator combination.
Caching also changes the bill. Hydrogen supports implicit caching, with a 95% discount on cached input tokens when prompts of at least 4,096 tokens get cache hits, but it does not support explicit context caching. Developers migrating an existing application must account for that change, as well as billing for thinking tokens and internal grounding instructions and function declarations. Google Search grounding remains billed per query.
Artificial Analysis recorded 304M output tokens for Hydrogen at xhigh, versus a similar-price-tier median of 72.8M. These evaluation totals do not measure Orchestrator or cost per coding task. Source: Gemini 4 Hydrogen (Xhigh) Intelligence, Performance & Price Analysis.
Google reports a Terminal-Bench 4.1 score of 57% for Hydrogen, against 18.2% for Gemini 3.8 Flash. Its model card also shows uneven results against competing models:
| Benchmark and reported measure | Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| SWE-sweep, subtask-weighted reward | 13.5% | 3.1% | 5.8% |
| Terminal-Bench 4.1, cheating-adjusted accuracy | 57.0% | 63.9% | Not reported |
| Agents’ Last Exam v1.1, rubric score | 71.6% | 60.4% | 69.8% |
| Terminal-Bench-Science 0.1, average reward | 51.4% | 59.9% | 58.3% |
Among these three models, Hydrogen leads the displayed repository-debugging and professional-workflow scores, while trailing Sonnet on Terminal-Bench and both rivals on scientific terminal tasks. Google's table also puts Hydrogen behind Sonnet on OSWorld-2.1 computer use: 77.6% against 83.9%, taking the best of three runs on an 82-task offline subset. The launch results support task-specific gains rather than a general competitive lead.
The evaluation methodology matters. Google generally reports the highest thinking settings, with stated exceptions. It computed Hydrogen’s Terminal-Bench results with a mini-swe-agent vision harness, averaging five attempts per task at high effort; the Sonnet comparison for Terminal-Bench 4.1 comes from the public leaderboard. Agents' Last Exam used the 122-task October 17 version, without the October 26 update, and a five-hour window for Google's runs; Terra's result came from the public leaderboard. Non-Gemini results generally come from providers’ reported figures unless otherwise specified. Google also says it used Terminal-Bench 4.0 for cost-versus-performance comparisons because competitor data across thinking levels was unavailable for 4.1. Harnesses, task versions and reasoning settings matter when translating benchmark scores into a purchasing decision.
Hydrogen is available on free and paid tiers in AI Studio and the Gemini API, and becomes the default model for the Antigravity agent in Gemini Managed Agents, Google's service for agents running in hosted sandbox environments. In the Antigravity development product, Hydrogen is rolling out to Google AI Pro and Ultra subscribers; Orchestrator initially goes to Ultra subscribers. Enterprises can access both through Gemini for business.
Hydrogen is proprietary, with no publicly available model weights, according to the independent analysis. Developers can buy API access and choose their workflow, or use Google's bundled environment to manage model handoffs. The SDK preserves an option for explicit role assignment, while Orchestrator puts the switching decision inside Google’s product.
The rollout also reaches the Gemini app, Search and eligible Google Sheets subscriptions. Free Gemini app users receive Hydrogen through automatic selection when appropriate for their prompt. In Search, Google says Hydrogen improves interpretation of complex questions, visual responses and query fan-out—the simultaneous searches across subtopics used to find relevant links. A report on the Search rollout contrasts Hydrogen’s global availability for all users with Carbon’s earlier availability in AI Mode for paid subscribers. Those are distribution choices; the launch evidence does not quantify their effect on website traffic.
Google’s illustrative AI Mode comparison shows Boston neighborhood responses from Gemini 3.5 Flash and Gemini 4 Hydrogen. Source: Google. Source: Introducing Gemini 4 Hydrogen.
Migration requires more than changing a model name for some users. Hydrogen shares a single 1,048,576-token budget across input, thinking and output. Its synchronous output cap is 65,536 tokens per hop; longer generation requires background execution or continuation. Developers must also preserve conversation state and thought signatures, remove explicit-cache requests and account for ignored custom sampling settings. Those constraints affect whether long-running agents complete reliably.
Google's safety announcement reports a 96.7% honesty rate on an internal agent evaluation. A different test in the model card records an increase of 5.57 percentage points in unjustified refusals versus Gemini 3.8 Flash, particularly affecting multimodal prompts, even as image-to-text content safety improves. These content-safety results come from automated evaluations, not human evaluation or red teaming. The card also warns of hallucinations, occasional slowness or timeouts, and a March 2026 knowledge cutoff.
Google says targeted testing found Hydrogen below its critical capability levels for biological-information and cybersecurity risks, while reaching alert thresholds in both domains. It did not require the full suite of critical capability evaluations under its Frontier Safety Framework because its assessment found no material increase in general capability over Argon. The full Safety & Security Report is still promised for the coming days.
For developers, the useful comparison is between Hydrogen alone at the lowest adequate thinking level, Carbon alone, and Orchestrator on the same work. Completion quality, elapsed time, retries, quota consumption and the total bill need to be measured together. API prices and subscription quotas describe different charging arrangements; the launch figures do not convert one into the other.
Google's benchmark-version discrepancy needs clarification, and the promised safety report should supply the fuller testing record. Until then, the evidence supports a new way to allocate reasoning and execution, with promising company-reported savings. Whether those savings survive a particular team's codebase, verification requirements and failure recovery remains the deployment test.
Japan is considering safety and reporting duties backed by enforceable orders for advanced AI. The proposal would strengthen its nonbinding framework, but capability thresholds, covered businesses and penalties remain unresolved.
Google says its new fast AI model is reaching Search users worldwide, improving complex answers and source discovery. The rollout expands access to Gemini 4, but its effect on answer quality and website traffic remains unmeasured.
Four nonprofits are asking a California court to stop five AI developers from building or deploying more capable systems. The case seeks enforceable limits even as OpenAI has previously reported a training pause and is launching defensive AI tools.
Expo Simulators gives coding agents remote iOS and Android environments with recordings for reviewers. The launch extends Expo's cloud development services into verification, with per-minute charges and shared capacity limits.
Kleiner Perkins and Clearlake Capital led Mirendil’s $1 billion Series A. Its agent-led model-building experiment points to smaller research teams, but company benchmarks, substantial compute costs and a private beta qualify the promise of wider access.
Mount Thor raised $45 million in equity, alongside separate hardware financing, to expand Apple infrastructure for AI agents and model inference. Existing Mac clouds and workload-specific performance limits put its demand and cost claims to the test.
Salesforce Ventures led a Series C that values Q-CTRL above $1 billion. The funding backs GPS-free navigation and quantum computing software, while undisclosed revenue and contested performance benchmarks leave the commercial case open.
Adobe demonstrated 10 experimental creative tools and opened Deep Brush and Hands On for testing. The lineup explores more direct control over AI edits, but existing alternatives and unproven production performance limit claims of novelty.
House Foreign Affairs Committee Chairman Brian Mast says outgoing Nvidia chief Jensen Huang has agreed to questioning, with a date still under negotiation. Speaker Mike Johnson says support for limiting chip exports to China is growing as lawmakers examine allegations involving Huang and a cloud partner.
Leaked correspondence and a reconstructed publishing program expose John Mark Dougan’s use of AI to stock imitation news sites with pro-Russian material. The records illuminate funding and working methods, while his claims of influencing other AI systems remain unproven.
William Blair raised AMD to Outperform as server-chip prospects and valuation improved. AMD already expects revenue growth above 80% in 2027, but customer incentives, custom accelerators and weaker quarterly cash generation complicate the investment case.
Apple’s $299 first smart display earns praise for fast home controls and useful kitchen features, but reviewers found gaps in how it keeps family members’ messages separate. Its strongest advantages also depend on owning compatible Apple devices and accessories.
Donald Trump proposed Apple executive chairman Tim Cook to succeed Jensen Huang at Nvidia. Cook’s interest and the seriousness of the suggestion remain unclear, while finance chief Colette Kress has taken interim charge.
Norwest’s largest fund is a third bigger than its predecessor, giving the Wells Fargo-backed investor more capital for AI alongside healthcare, consumer and industrial businesses. The firm has not disclosed how much will go to AI.
Nebius raised its 2026 capital spending forecast to about $27 billion while steering revenue to the low end of guidance. Better customer prepayments support expansion, but delayed capacity and a $325 million non-recurring contribution complicate its quarterly revenue beat.
Megaspeed founder Alice Huang is reportedly living in Beijing, but her activities and the espionage allegations remain unresolved. Jensen Huang’s resignation changes Nvidia’s leadership without settling questions about her business connections and access to its chips.
Singapore says it barred former Megaspeed chief Alice Huang after she gave inaccurate information about her role, while finding insufficient evidence against her over server declarations. The wider investigation still depends on tracing shipments beyond Malaysia.
Nvidia’s longtime finance chief takes over from its co-founder as the board promises continuity, searches for a permanent CEO and faces unresolved questions over security and executive conduct.
Chicago Public Schools is seeking a way to preserve Google Search without its generative AI features as parents, educators and students press for stronger district rules. Google’s planned access changes turn an AI safety dispute into a test of who controls classroom technology.
Warby Parker scheduled November 12 preorders and holiday deliveries for its first Android XR glasses. Google’s backing combines development funding with a conditional equity investment, while the retailer expects only a modest initial sales contribution.