Gemini 4 Hydrogen beats Argon on one index at under half the price

Google has released Gemini 4 Hydrogen, a faster and cheaper model in its Gemini 4 family. According to the official announcement, it scores 51 on the Artificial Analysis Intelligence Index at its default setting, ahead of the larger Gemini 4 Argon at 50.

Koray Kavukcuoglu, Google DeepMind’s SVP and Chief AI Architect, wrote that Hydrogen is meant for “bringing cutting-edge intelligence to everyone”. Google positions it below Gemini 4 Carbon and Argon, which it describes as the frontier models for deep, long-horizon reasoning.

Flat pricing and a broad rollout from day one

Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Google says the price is the same across the whole 1 million-token context window, so long prompts do not move into a more expensive tier. The model can also produce up to 1 million output tokens, according to the post.

Availability is unusually wide for a launch. Google says the rollout is global and starts today, covering every plan level of Google AI Studio and the Gemini API, plus Gemini for business, the Gemini app and Google Search. Paid Google AI Pro and Ultra subscribers can use it in Google Antigravity. Developers on the paid tier also get it in Android Studio. Gemini in Google Sheets gets it first for Pro and Ultra subscribers, and for eligible business plans.

It has four thinking levels: low, medium, high (the default) and a new xhigh tier. On the Artificial Analysis index, Google says xhigh reaches 54. The company says high finishes tasks 2.5 times faster than Gemini 3.8 Flash, and that Hydrogen’s Terminal-Bench 4.1 score is 57% against 18.2% for that model.

Google also introduced Orchestrator in Antigravity. It switches between models such as Carbon and Hydrogen inside a single session. On Terminal-Bench 4.0, Google says it reached near-Carbon success rates while using 24% less quota and finishing 19% faster. For parallel agents, Carbon can coordinate while Hydrogen runs the worker and verifier agents.

Priced at 37.5% of Argon’s list rate

The headline comparison is with Argon, and the numbers hold up on the benchmark site’s own pages. Artificial Analysis lists Argon at 50 on version 5.0.1 of its index, with a standard price of $4 per million input tokens and $20 per million output tokens, and a 1M-token context window. Hydrogen’s $1.50 and $7.50 are therefore 37.5% of Argon’s list price on both sides of the meter. That makes the 51-to-50 gap more notable than it looks: Google is offering a score one point higher for well under half the token cost.

The comparison needs care. Search results show Argon was introduced with a temporary 50% discount, at $2 and $10, and the end date of that promotion is reported inconsistently by secondary sources. Against the discounted rate, Hydrogen is still cheaper, but by a smaller margin. Token price is also not task cost. Reasoning models differ in how many output tokens they spend, which is why Google stresses fewer output tokens across thinking levels, though the post gives no figure for that.

Against Hydrogen’s predecessor the picture is different. Several price trackers list Gemini 3.8 Flash at an introductory $0.75 and $3.75 per million tokens, with a reported step up to $1.50 and $7.50 on January 1, 2027. These are third-party listings that were not checked against Google’s pricing page. If they are right, Hydrogen’s price matches what 3.8 Flash was reported to cost after the promotion, while Google claims much higher scores and speed.

What the benchmark claims leave unproven

Almost every figure in the post is Google’s own. That includes Hydrogen’s #5 rank on arena.ai’s Agent Arena, its 58.8% on HLE-Diamond, 52.3% on the Vals Tax Agent Bench, 55.8% on the Vals Legal Research Bench and 53.4% on the new Multimodal Knowledge Agent benchmark. Google reports these as leading or tying within its class, but the post does not name the competing models in the text or say how each was configured. The MKA benchmark is new, and the post does not say who built it.

The safety figures are also self-reported: up to 81% less overreach in agentic tasks, a 96.7% score on Google’s internal honesty evaluation, and false refusals under 0.4% on HealthBench Pro. Google points to the Hydrogen model card for details. The speed claims for partner workloads, 2x to 2.8x faster with 12% to 23% fewer tokens than 3.8 Flash, come from early testing by unnamed partners.

For Search, Google says Hydrogen improves understanding of multi-part questions and typos, and widens the range of links shown through its query fan-out technique. A separate Search post by Robby Stein calls it the biggest model upgrade to Search yet, applies it to all supported AI Mode languages, and gives no usage figures or per-country schedule.

For developers, the practical test is the one the post itself sets: whether a cheaper execution model, paired with Carbon or Argon for planning, holds up in their own repositories. Independent runs of the Artificial Analysis index and of coding benchmarks should show that within days.