⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 6 min read

Gemini 4 Hydrogen vs Argon: Every Published Benchmark Side by Side, and What It Means

Google-reported benchmark results for Gemini 4 Hydrogen and Argon, read with the settings that move them.

Two gold bars rising on a benchmark leaderboard grid over dark navy

Gemini 4 Hydrogen vs Argon pits Google's fast default model against its efficient frontier reasoner, and on Google's own published numbers Hydrogen leads on agentic coding while Argon leads on chart reading.

Book a free SEO strategy session →

GoldieBench has not run either model yet, so this page sticks to vendor-reported results and says clearly where they come from.

To see how we apply model testing to SEO, book a free SEO strategy session.

Sources are the DeepMind Hydrogen and Carbon model cards and Google's changelog.

The scoreboard Google published

BenchmarkHydrogenArgonCarbonCategory
SWE-sweep13.5%4.3%9.6%Agentic coding
OSWorld-2.1 (offline subset)77.6%74.7%81.5%Computer use
Chartography66.6%71.6%72.5%Multimodal

Hydrogen's numbers come from its November card; Argon's and Carbon's come from the Carbon card dated October 2026.

Cross-card comparisons carry a caveat: methods and dates differ slightly, so small gaps can move.

What the cards compare against

Google benchmarked Hydrogen against Claude Sonnet 5.5 and GPT-6.1 Terra rather than against Argon.

On its card, Hydrogen posts 71.6% on Agents' Last Exam v1.1 and 58.8% on HLE-Diamond, ahead of both of those rivals on each.

Argon's own card compares it with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 instead.

Settings that change results

  • Thinking level: Hydrogen defaults to high with an xhigh option; Argon defaults to medium.
  • Output budget: both share a 1,048,576-token window and stop at 65,536 per synchronous response.
  • Non-streaming hops: Hydrogen up to 65,536, Argon 32,768 over REST.
  • Sampling: Hydrogen ignores in-range temperature, top_p and top_k and rejects out-of-range values.

A benchmark run that leaves these at default is comparing different effort settings, which is why GoldieBench fixes one prompt and one setting per model.

How we would test them on GoldieBench

The bench gives every model the same fixed prompt, one HTML file out, scored 0-10 on whether it ran, how close it hit the brief and how it looked.

Book a free SEO strategy session →

For a fair Hydrogen vs Argon run, both would use matched thinking levels, one shot, no retries.

Until those runs exist, the vendor table above is the best public evidence.

Model choice by job

JobLean towardsEvidence
Fast iterative codingHydrogenHigher SWE-sweep on Google's cards
Computer-use agentsHydrogenHigher OSWorld-2.1 offline score
Charts and figuresArgonHigher Chartography score
Long, hard reasoning with less spend than CarbonArgonGoogle's positioning

Related models already on the bench

GoldieBench currently ranks Claude Opus 5.5 at an average of 7.57 across 50 one-shot builds, and the newest Gemini on the leaderboard is Gemini 3.6 Flash.

Gemini 4 Hydrogen and Argon are not on the leaderboard yet.

Argon's full card, in brief

Benchmark (Argon card, Oct 2026)ArgonClaude Opus 5.5GPT-6 Astra
DeepSWE v1.177.9%74.2%74.1%
Vals Index68.9%67.0%63.1%
AutomationBench51.3%42.5%41.4%
GraphWalks 256K to 1M (BFS F1)84.2%66.8%71.8%
LVBench91.7%83.7%87.5%

These rows are taken from the Argon model card, where Google lines Argon up against Anthropic and OpenAI models.

They suggest Argon's edge is long context and video, not raw coding speed.

Hydrogen's card, in brief

Benchmark (Hydrogen card, Nov 2026)HydrogenClaude Sonnet 5.5GPT-6.1 Terra
SWE-sweep13.5%3.1%5.8%
Agents' Last Exam v1.171.6%60.4%69.8%
OSWorld-2.1 offline77.6%83.9%74.2%
HLE-Diamond58.8%53.3%47.3%

On the Hydrogen model card, Sonnet 5.5 beats Hydrogen on OSWorld-2.1, so Hydrogen is not the computer-use leader across all models, only ahead of Argon.

Speed and cost signals

Google labels Hydrogen fast, Argon moderate and Carbon deliberate.

Hydrogen's list price per token is below Argon's, and Argon's introductory rate expires on 30 November 2026.

Book a free SEO strategy session →

For a one-shot bench like ours, speed shows up as build time per task, which is one of the numbers we publish next to each score.

What we expect to check first

  • Whether Hydrogen's SWE-sweep lead turns into cleaner one-shot HTML builds.
  • Whether Argon's long-context strength helps on tasks with large briefs.
  • How much xhigh on Hydrogen changes quality against its time cost.

How to read a vendor card

Every lab picks the rivals and benchmarks that flatter its model, and Google is no different.

Hydrogen's card compares it with Sonnet 5.5 and GPT-6.1 Terra, models in a similar speed class, rather than with its own bigger siblings.

Argon's card compares it with the frontier models of its launch month.

Cross-reading the two cards is the only way to put Hydrogen and Argon in one table, and that is why the caveats above matter.

Thinking levels as a hidden variable

Scores on a model card are usually run at the model's best setting, not its default.

Hydrogen's best setting is xhigh, which Google says uses the most tokens and has the slowest first token.

Argon's best is high.

If you run either at its default and get a lower score than the card, that is expected, not a bug.

Multi-model setups

Google's Hydrogen guide describes using Hydrogen as the worker and verifier alongside Carbon or Argon.

In one session, Carbon or Argon can plan the interfaces or clear a hard roadblock, then hand code edits, parallel subtasks and test runs to Hydrogen.

GoldieBench's Hermes MoA entry already tests a mixture-of-agents setup on the bench, so this kind of pairing is something we can measure, not just describe.

What would change our view

If a one-shot GoldieBench run showed Argon producing clearly better builds than Hydrogen at the same thinking level, the vendor table would matter less.

If Hydrogen matched Argon at medium, the case for Argon on build tasks would mostly disappear.

Either way, the result will land on the leaderboard with a playable demo next to every score.

Release timing

Argon has had a month in the wild; Hydrogen only days.

Early numbers for any model tend to move as providers tune serving and as independent benchmarks catch up, so expect this table to change.

We will update the page when bench runs land.

Verdict for now

On published numbers, Hydrogen is the stronger general agent model and Argon the stronger chart reader.

For tested agent setups, the AI Profit Boardroom has the trainings and the Agent OS.

Book a free SEO strategy session →

Match the thinking levels before you trust any chart and Gemini 4 Hydrogen vs Argon becomes a fair fight; for data-led SEO, book a free SEO strategy session.

FAQ

Has GoldieBench tested Gemini 4 Hydrogen or Argon?

Not yet. This page uses only Google-published model card results until bench runs are done.

Which scores higher on SWE-sweep?

Hydrogen, at 13.5% on its card, against 4.3% for Argon on the Carbon card.

Which is better at reading charts?

Argon, at 71.6% on Chartography against Hydrogen's 66.6%.

Why compare across two model cards?

Google did not publish Hydrogen and Argon in one table, so the rows combine the Hydrogen card and the Carbon card, which lists Argon.

Is Hydrogen better than Claude Sonnet 5.5?

On Google's card Hydrogen leads Sonnet 5.5 on SWE-sweep, Agents' Last Exam and HLE-Diamond, but Sonnet 5.5 scores higher on OSWorld-2.1 offline.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week