# Google's Gemini 4 Hydrogen scores 54 on the Artificial Analysis index, ahead of the larger Argon, at a lower list price

Hydrogen lists at a reported $1.50 per million input tokens against Argon's $2 introductory rate, which doubles after November 30. Google is also putting the model into Search worldwide.

- Published: 2026-11-13T06:16:33.119Z
- Canonical: https://polylog.news/ai/2026-11-13/google-s-gemini-4-hydrogen-scores-54-on-the-artificial-analy
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Google DeepMind](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/), [Polylog editors](https://polylog.news), [Google AI Blog](https://blog.google/products-and-platforms/products/search/gemini-4-hydrogen-search-upgrade/)

Google [released Gemini 4 Hydrogen](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) on November 12 and positioned it for coding, agents and long multi-step automation. On the Artificial Analysis Intelligence Index, Hydrogen scores 54 at its "xhigh" thinking setting, [against 50 for Gemini 4 Argon](https://t.me/ai_machinelearning_big_data/11348). [The Decoder](https://the-decoder.com/googles-gemini-4-hydrogen-outscores-the-bigger-argon-just-four-weeks-after-its-launch/) notes that Argon is the larger model and launched about four weeks earlier. At the default "high" setting Hydrogen scores 51. Gemini 4 Carbon, Google's top tier, scores 60.

In Google's own ten-benchmark table, Hydrogen leads Claude Sonnet 5.5 and GPT-6.1 Terra on seven tests, [according to a breakdown of the table](https://help.apiyi.com/en/gemini-4-hydrogen-official-benchmark-10-tests-7-wins-en.html). The reported results include 71.6% on Agents' Last Exam v1.1 and 13.5% on SWE-sweep. Sonnet 5.5 leads on the three tests closest to agentic shell work: Terminal-Bench 4.1, OSWorld-2.1 computer use and Terminal-Bench-Science. One report puts Terminal-Bench 4.1 at [57% for Hydrogen against 63.9% for Sonnet 5.5](https://the-decoder.com/googles-gemini-4-hydrogen-outscores-the-bigger-argon-just-four-weeks-after-its-launch/). On Artificial Analysis's AA-Omniscience test, Hydrogen fabricates answers 8% of the time.

Price is the more consequential number. Third-party trackers list Hydrogen at [$1.50 per million input tokens and $7.50 per million output tokens](https://www.benchleader.com/models/gemini-4-hydrogen), with a 1-million-token context window and a [2.5x speed claim attributed to Google](https://shattered.io/gemini-4-hydrogen-agentic-workflows-1m-context-1-50-input-2026/). Argon's introductory price is $2 and $10, and it [rises to $4 and $20 after November 30](https://fourweekmba.com/ai-gemini-4-argon-goes-global-price-doubles-after-nov-30/). One analysis puts Hydrogen's cost per task at [$1.87 against $7.77 for Gemini 3.8 Flash](https://aimlapi.com/blog/gemini-4-hydrogen-vs-3-8-flash-vs-argon), with a higher score. Another reports that Hydrogen [matched a pricier model on a Zapier automation test at half the cost](https://xenospectrum.com/en/gemini-hydrogen-benchmark-comparison/).

Google is also [bringing Hydrogen to Search globally](https://blog.google/products-and-platforms/products/search/gemini-4-hydrogen-search-upgrade/). The rollout gives the model a large default audience before any developer chooses it.

Nearly all of these figures come from Google or from aggregators that repeat Google's table. The Artificial Analysis index scores are the main independent measurement, and aggregators disagree on some Terminal-Bench figures and on which index version they use.

## What this means

Google is using a smaller model to undercut its own larger model and its rivals' mid-tier models on price per task. Anthropic's Sonnet tier and OpenAI's GPT-6.1 Terra are exposed through price, and API resellers that earn margin on mid-tier traffic are exposed as well. Google also serves Search on its own Tensor Processing Units, so it spreads inference cost across a product with more than a billion users. That is a cost structure labs that rent GPUs do not have. Most coverage reads a 54 on a composite index as a significant increase in capability. The Terminal-Bench results tell a narrower story: Anthropic still leads on agentic shell work, the use case coding-agent buyers pay most for. The desk's forward read is that within weeks a rival lab will cut mid-tier prices or launch a cheaper tier, and that Argon's price increase after November 30 will move developer volume toward Hydrogen rather than keep it on Argon. This depends on independent agentic evaluations confirming Google's table. The read would be wrong if Vals AI, LiveBench or Artificial Analysis agent tests place Hydrogen well below Sonnet 5.5 on coding tasks.

## What to watch

- Whether Google holds Argon's price increase on November 30 or extends the introductory rate. Holding it would mean Google intends Hydrogen to be its main developer model.
- Independent Terminal-Bench and SWE-style results for Hydrogen from evaluators other than Google. These will show whether the seven-of-ten table holds on agentic coding.
- Pricing responses from Anthropic and OpenAI on their mid-tier models over the coming weeks.