Gemini 4 Carbon vs Gemini 4 Argon

Verdict

Gemini 4 Carbon scores higher — 60 vs 50, both at their highest reasoning tier (high). Gemini 4 Argon costs 28% more per task ($9.95 vs $7.76). Per index point: $0.1293 for Gemini 4 Carbon vs $0.1990 for Gemini 4 Argon.

Gemini 4 Carbon ranks #61 of 73 on cost per task among closed-source models above $1 per million tokens (blended 3:1) that AA prices; the class holds 237 models. Gemini 4 Argon ranks #65 of 73 on cost per task in that same class of 237.

Our pick: Gemini 4 Carbon. The price sentences above use it as the baseline; scope: API access.

• Gemini 4 Carbon — Take Gemini 4 Carbon on both the bill and the clock: 60 vs 50 index points, $7.76 vs $9.95 per task ($0.22 per extra point), 25.7 vs 30.0 minutes.
• Gemini 4 Argon — Gemini 4 Carbon leads on four of the nine axes on its own merits (Intelligence Index 60 vs 50, cost per task $7.76 vs $9.95, minutes per task 25.7 vs 30.0, input modalities) and the other four are level; because it wins both per-task cost and the Intelligence Index, it also wins cost per index point ($0.1293 vs $0.1990), a ratio of the two. No case for Gemini 4 Argon can be derived from the data on these pages.

Side by side

Where each wins

Gemini 4 Carbon

• Intelligence Index: 60 vs 50
• Cost per task (USD): $7.76 vs $9.95
• Minutes per task: 25.7 min vs 30.0 min
• Input modalities: text, image, speech, video vs text, image
• Cost per index point: $0.1293 vs $0.1990

Gemini 4 Argon

• Level on Input $/M tokens ($4.00)
• Level on Output $/M tokens ($20.00)
• Level on Cache discount (95%)
• Level on Context window (1M)
• Lower on Intelligence Index; more expensive per task; slower per task; narrower on Input modalities (text, image vs text, image, speech, video); worse on cost per index point

Sources and conflicts

• OpenRouter lists google/gemini-4-carbon (exact match). Input price: AA $4 per million tokens vs OpenRouter $2 — a 2.0× gap. This page uses AA's figure throughout and does not average the two.
• OpenRouter lists google/gemini-4-argon (exact match). Input price: AA $4 per million tokens vs OpenRouter $2 — a 2.0× gap. This page uses AA's figure throughout and does not average the two. Modalities: AA records text and image; OpenRouter's route declares image, speech, text, video.
• Prices and input modalities come from Artificial Analysis only; the second source is used to cross-check, never to union or average.
• A differing modality string is not a conflict: AA records what the model accepts, a provider records what its own route declares. Different objects, both reported here.

Data

Data as of 2026-11-12 · Artificial Analysis Intelligence Index v5.0.1, highest reasoning effort per model. Prices from the same pages.
Every cell is tied to a verbatim excerpt of that model's own Artificial Analysis page.
N/A (AA: Unknown) = AA publishes no value in that cell. It is not 0, and it is not 'unreleased'.
'not applicable (AA prints `-`)' = AA lists no per-token price for this model, so the cell does not exist. It is not the same as N/A and not the same as a withheld figure.
'not computable (…)' = the cell is derived from two others and one of them has no figure. The bracket names which one — it is not a value, and it is not the same as N/A.
'not printed on its AA page' = the page lists prices but withholds that figure; it is not treated as equal to anything.
'unverified' = we have not read that cell on that page. It is not a value and not a tie.
Token rows are AA's Verbosity block: input and output totals for the whole Intelligence Index run. 'Output/input ratio' is AA's own label and means output / input — it is not output's share of the total.
AA publishes no measured cache-hit rate for either model (it prints a fixed 7:2:1 cache-hit/input/output blend instead), so no cell records one.
Token blocks: AA prints run-level token totals in its Verbosity block and the per-task bill in its cost block; the two use different denominators, so each is cited where it belongs and neither is derived from the other.