Gemini 4 Hydrogen: $1.50 in, $7.50 out per million tokens

Google put a Gemini 4 model on every free surface it owns today, and priced it below everything else in the family. The benchmark is fine. The price is the story.

Google launched Gemini 4 Hydrogen today at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1 million-token context window with no long-context surcharge — the cheapest tier in the Gemini 4 line, and 25% below the introductory rate Google is still charging for Gemini 4 Argon, according to the launch post. It is available the same day in AI Studio, the Gemini API, Android Studio's paid tier, Gemini for business, the Gemini app, Google Search for every user globally, and Google Antigravity for paying AI Pro and Ultra subscribers. Free users in the Gemini app only reach it through auto model selection; Search serves it to everyone.

The positioning is explicit and worth reading carefully, because it reframes what a "model launch" is now. Google calls Carbon and Argon its frontier models — the ones you use for "deep, long-horizon reasoning on your hardest problems" — and describes Hydrogen as an execution model: writing code, running tests, verifying changes, calling tools, doing the loop work. Google's own framing is that teams keep a frontier model for architectural judgment and hand the repetitive execution to something faster and cheaper. Hydrogen can stand alone, or run underneath Carbon and Argon as the engine.

That tier split is not unique to Google. Every major lab now sells a reasoning flagship and a cheap workhorse, and the workhorse is where the token volume actually lives. What is new is the price ratio inside one family. Carbon, generally available since last week, lists at $4 per million input and $20 per million output once its introductory rate ends — we covered its arrival on top of the index. Hydrogen undercuts that list price by 62% on input and 62% on output, and comes in 25% under Argon's $2 and $10 introductory pricing, which is the rate most developers have actually been paying this month.

The capability gap is narrower than the price gap. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high thinking level — one point above Argon's 50 — and 54 at a new "xhigh" setting above Google's previous top tier. Carbon sits at 59 on the same index. So Google is selling a model that roughly matches its previous flagship on a broad composite, with the top of its range within five points of the current flagship, for a quarter to a third of the price on input. That is a strange thing for a vendor to do to its own product ladder, and it is the strongest signal yet that per-token pricing in the sub-flagship tier is in free fall.

Against the model Hydrogen actually replaces, the jump is large. Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at the default high setting, more than triples Terminal-Bench 4.1 scores (57% versus 18.2%), and cuts fabricated claims by more than five times. On partner workflows across complex graphs, SQL and data-heavy enterprise jobs, Google reports Hydrogen running 2 to 2.8 times faster while using 12% to 23% fewer tokens than 3.8 Flash. Those are vendor numbers, but the direction is consistent with the benchmark gap, and token reduction matters more than latency here: a cheap model that writes three times as much output is not cheap.

The most interesting engineering is in Antigravity, not the model. Google shipped an Orchestrator that switches between Carbon and Hydrogen inside a single session instead of running one model start to finish. On Terminal-Bench 4.0, Google says Orchestrator reached near-Carbon task success rates while consuming 24% less quota and finishing 19% faster — and that parallel agent workflows can put Carbon in the coordinator role with Hydrogen running workers and verifiers in isolated workspaces. That is a routing problem dressed as a product feature, and it is where the real cost battle will be fought: if a scheduler can send easy steps to the cheap model and hard ones to the expensive one, the blended bill drops without anyone choosing a model.

The benchmark sheet is broad but honest about its own shape. Hydrogen ranks fifth on Agent Arena, leads its class on HLE-Diamond at 58.8%, ranks first overall on the Vals Tax Agent Bench at 52.3%, ties for first on Vals' Legal Research Bench at 55.8%, scores 53.4% on Google's new Multimodal Knowledge Agent benchmark for end-to-end work over PDFs, spreadsheets and recorded video, and 71.6% on Agents' Last Exam v1.1. It takes text, image, speech and video in, emits text, and pairs 1 million tokens of input with up to 1 million of output — a ceiling Google only recently raised on Argon, and now the default across the family.

The safety section reads like a direct answer to the agent-reliability complaints of the past year. Google claims Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, cuts misaligned actions and task shortcutting across tool-use and coding evaluations by 12 times, is up to twice as likely to stop when a task is impossible rather than route around a permission boundary, and scores 96.7% on an internal agent honesty evaluation. Against indirect prompt injection it claims to be more than 2.3 times more resilient on the Gray Swan benchmark. Keep the 99.9% chain-of-thought legibility claim in the vendor-claim column until someone outside audits 200,000 turns, but the permission-boundary number is the one that matters for anyone running agents against real systems.

The counter-argument comes from the same independent benchmark Google cites for its headline. Artificial Analysis measures Hydrogen at 41.1 seconds to first answer token at the xhigh setting, against a median of 13.45 seconds for models in its price class, and 198 output tokens per second — slow enough that its 13 minutes per index task is slightly worse than the class median despite the "fast" positioning, per its model page. Worse for the cost story: it generated 304 million output tokens across the index, placing it 76th of 77 on verbosity, and cost $3.16 per task. Cheap per token, mid-priced per task. Hydrogen's reasoning tokens are where the bill hides, and the xhigh setting that produces the 54 score is exactly the one that took 41 seconds to start answering. Anyone budgeting an agent fleet should price the task, not the token.

The pattern to watch is the price floor, not this model. Google has now shipped three Gemini 4 tiers in five weeks, each with an introductory rate it later doubles, and each one cheaper than the last on input. The Search rollout is the part with real scale behind it: Google is serving Hydrogen to every Search user worldwide and calling it the biggest model improvement to Search yet, which quietly moves the highest-volume AI workload on earth onto a Gemini 4 model. If Search inference is now running on $1.50-per-million input, the unit economics of the whole consumer AI market just moved. Whether the public rate survives past the intro window — Argon's and Carbon's both doubled — is the question every procurement team should be asking this week.

United Airlines will block third-party AI agents from booking flights and managing reservations, citing credential-sharing risk and federal aviation rules, the airline said in its own statement. United warns that handing account credentials to a personal AI assistant can expose sensitive data and produce unintended purchases, account changes or drained MileagePlus miles, and it says agents can break rules that require a human to complete check-in and confirm a bag holds no banned substances. The carrier is restricting access both inside its own apps and through travel agencies and aggregators, with changes to booking, ticketing and distribution policies — the operational detail that makes this more than a press release, since it puts the burden on intermediaries to detect and block agents. It follows Bank of America's warning last week that customers should not share passwords with AI assistants, and a wider pattern of consumer platforms drawing a line: Amazon and eBay already restrict agent access, and Resy banned diners whose agents grabbed tables. The airlines are not wrong about the liability — a check-in attestation signed by a bot is a compliance problem, not a UX preference — but "share your password with us or use our app" is a defense that only holds until a booking agent gets an official delegation standard to authenticate against.

If a cheap execution model still burns reasoning tokens, is the per-token price you see on the pricing page the one you should budget against? Write to us via our contact page.