Skip to content
← All posts
Live comparison·7 min read

Gemini 4 Argon vs Carbon vs Hydrogen: Thinking Levels Compared Live

Compare Argon, Carbon and Hydrogen on the same Driftwood brief, then compare each model at its supported thinking levels. Real outputs, measured tokens and per-run cost.


Argon, Carbon and Hydrogen are all in the Gemini 4 arena. They build the same fictional product: Driftwood, a service that turns travel photos into hardcover books. This comparison separates two questions: what changes between models at medium, and what changes when the thinking level changes within one model.

Published coverage, not a leaderboard

10 of 10 supported model-and-level combinations have been tested. Published Driftwood samples: 8. Each published page is one completed sample (n=1), preserved without code edits. Output-limit failures recorded separately: 2. Pending settings are not results or scores. The live coverage is also shown on the Gemini 4 hub.

Which thinking levels are supported?

ModelSupported thinking levelsSetting in the original runs
Gemini 4 Argonlow, medium, highExplicit medium
Gemini 4 Carbonlow, medium, highExplicit medium
Gemini 4 Hydrogenlow, medium, high, xhighExplicit medium

These are the settings listed in Google's thinking-level documentation. Hydrogen defaults to high, but our original Hydrogen run explicitly requested medium. A level is a requested setting, not a fixed number of reasoning tokens. The names are not a standardized amount of computation across different model families.

First compare the three models at medium

The three medium outputs make visibly different design choices. Argon uses a terracotta book cover and the headline "Your trips deserve better than a forgotten cloud folder." Carbon opens a Kyoto field-notes spread under "Your camera roll is a graveyard of great trips." Hydrogen presents a linen Kyoto, Autumn cover with color swatches. These are observations about the actual pages, not a quality ranking.

Gemini 4 Argon · medium
Gemini 4 Carbon · medium
Argon medium and Carbon medium: the same original Driftwood brief. Open “Landing” in the arena to run both full-screen, go blind, and vote.
Gemini 4 Carbon · medium
Gemini 4 Hydrogen · medium
Carbon medium and Hydrogen medium: compare the same product at the same requested level. Open “Landing” in the arena to run both full-screen, go blind, and vote.

Open all three medium outputs together. Keep the model comparison separate from the thinking-level comparisons below: a more expensive model and a higher thinking setting are different variables.

Measured usage and cost by requested setting

ModelThinkingStatusBilled output tokensVisible output tokensReasoning tokensSample cost
ArgonlowPublished24,06324,063Not reported$0.120656
ArgonmediumPublished26,600Not recorded hereNot recorded here$0.133341
ArgonhighPublished38,75936,8981,861$0.194136
CarbonlowPublished29,83129,831Not reported$0.149496
CarbonmediumPublished35,099Not recorded hereNot recorded here$0.175836
CarbonhighPublished62,76049,66613,094$0.314141
HydrogenlowPublished19,70518,805900$0.074150
HydrogenmediumPublished25,42323,3462,077$0.095592
HydrogenhighOutput limit (MAX_TOKENS)65,53663,1752,361$0.246016
HydrogenxhighOutput limit (MAX_TOKENS)65,53655,8679,669$0.246016

Billed output includes reasoning. Reasoning tokens appear only when the provider reports them and that field has been audited; an absent count does not mean zero thinking. Cost is calculated from recorded request usage at the Flex price in force on its run date. It excludes failed or unresolved infrastructure attempts, so it is a sample cost, not the total project bill.

Hydrogen high: output limit reached

The Hydrogen high request ended with MAX_TOKENS at the shared 65,536-token cap: 63,175 visible-output tokens and 2,361 reasoning tokens. The recorded sample cost was $0.246016.

The provider reported MAX_TOKENS at the shared output cap. The archived response ends in unfinished CSS. Independent parsing confirmed one response and one candidate with incremental SSE chunks. We preserved the full failed attempt and have published no playable page from it. The raw response remains unchanged, and no quality reroll was made.

The exact requested setting was HIGH, on Flex. This is one observed output-limit failure, not evidence that every high-thinking run will fail.

Hydrogen xhigh: output limit reached

The Hydrogen xhigh request ended with MAX_TOKENS at the shared 65,536-token cap: 55,867 visible-output tokens and 9,669 reasoning tokens. The recorded sample cost was $0.246016.

The provider reported MAX_TOKENS at the shared output cap. The archived response ends in unfinished CSS. Independent parsing confirmed one response and one candidate with incremental SSE chunks. We preserved the full failed attempt and have published no playable page from it. The raw response remains unchanged, and no quality reroll was made.

The exact requested setting was XHIGH, on Flex. This is one observed output-limit failure, not evidence that every high-thinking run will fail.

Argon: low, medium, high

The Argon comparison keeps the Driftwood brief, model and output cap fixed. Only the requested thinking level changes. 3 of 3 supported settings have a published landing page in this snapshot.

Observed defects in Argon high: Hero paragraph and nearby trust text are clipped at the right edge. The hero paragraph right bound is 409.765625px in a 390px viewport; its left bound is 24px. The original HTML is preserved, and this result remains available for comparison.

Recorded costs for these individual samples: low: $0.120656; medium: $0.133341; high: $0.194136. Each cost belongs to one run. It does not establish a quality ranking or consistency across repeated runs.

Gemini 4 Argon · low
Gemini 4 Argon · high
Argon at low and high. Unedited outputs of the same original brief. Open “Landing” in the arena to run both full-screen, go blind, and vote.

Open every published Argon level side by side. Check the FAQ, narrow-screen navigation and main call to action as well as the first screen. A longer answer does not establish that those interactions work better.

Carbon: low, medium, high

The Carbon comparison keeps the Driftwood brief, model and output cap fixed. Only the requested thinking level changes. 3 of 3 supported settings have a published landing page in this snapshot.

Observed defects in Carbon high: The header menu button is partially clipped at the right edge on the 390px mobile view. The same clipping remains visible after navigation. The original HTML is preserved, and this result remains available for comparison.

Recorded costs for these individual samples: low: $0.149496; medium: $0.175836; high: $0.314141. Each cost belongs to one run. It does not establish a quality ranking or consistency across repeated runs.

Gemini 4 Carbon · low
Gemini 4 Carbon · high
Carbon at low and high. Unedited outputs of the same original brief. Open “Landing” in the arena to run both full-screen, go blind, and vote.

Open every published Carbon level side by side. Check the FAQ, narrow-screen navigation and main call to action as well as the first screen. A longer answer does not establish that those interactions work better.

Hydrogen: low, medium, high, xhigh

The Hydrogen comparison keeps the Driftwood brief, model and output cap fixed. Only the requested thinking level changes. 2 of 4 supported settings have a published landing page in this snapshot.

Recorded costs for these individual samples: low: $0.074150; medium: $0.095592. Each cost belongs to one run. It does not establish a quality ranking or consistency across repeated runs.

Gemini 4 Hydrogen · low
Gemini 4 Hydrogen · medium
Hydrogen at low and medium. Unedited outputs of the same original brief. Open “Landing” in the arena to run both full-screen, go blind, and vote.

Open every published Hydrogen level side by side. Check the FAQ, narrow-screen navigation and main call to action as well as the first screen. A longer answer does not establish that those interactions work better.

What makes this a controlled comparison?

  • The original Driftwood task brief and API system instructions stay fixed. Every setting receives the same required sections, offline HTML requirement and output format.
  • Every request explicitly records its model and thinkingLevel. maxOutputTokens stays at 65,536, including reasoning. Higher thinking can therefore leave less room for visible HTML within the same cap.
  • These runs use the Gemini API Flex tier with no tools, manual follow-up prompts or output edits. There are no reruns to obtain a prettier design. Infrastructure failures and any controlled infrastructure retries remain separate recorded attempts.
  • We do not turn wall-clock time into a generation-speed score. Flex can queue a request before it starts producing tokens.
  • One sample per setting cannot establish consistency or a statistical winner. Compare the real pages, including their broken or awkward interactions, before drawing a conclusion.

The Gemini 4 hub shows published coverage and model filters. The method page explains the shared task briefs and differences between API and CLI execution routes. Google AI Studio pricing documents the provider rates.

Frequently asked questions

Do all three Gemini 4 models support xhigh?

No. Argon and Carbon support low, medium and high. Hydrogen also supports xhigh. The comparison only uses settings supported by each model.

Does a higher thinking level guarantee a better page?

No. The setting changes the requested reasoning effort. It does not guarantee better design or working interactions. These are individual samples, not a statistical ranking.

Are all ten model-and-level results published?

10 of the ten supported combinations have been tested in this build, with 8 published Driftwood samples. Output-limit failures recorded separately: 2. The Gemini 4 hub lists each setting and distinguishes failed attempts from settings not yet published.

Don’t take the post’s word for it

The arena runs every model’s real output live. Pick a challenge, go blind, and cast a vote that counts in the public tally.

Open the arena