In this page snapshot, Gemini 4 Carbon's High, Medium, and Low tiers scored 60, 58, and 54, respectively, on the Intelligence Index. High scored best, but took the longest per index task (32.1 minutes), with a weighted cost of $5.48 per task.
Tasks this is suitable for assessing: Comparing the three reasoning tiers listed on the page in terms of their relative performance on the Artificial Analysis Intelligence Index, task duration, output speed, and index task cost.
Tasks this should not be generalized to: These results do not establish performance on all real-world business tasks, with other clients, or with different prompts. Per-task cost should not be treated as the total cost of running the full index.
Applicable model version: Gemini 4 Carbon. The page lists High, Medium, and Low separately.
Test environment or client: The Artificial Analysis model page. The page says the index evaluation is independently measured by Artificial Analysis. The specific calling endpoint is not identified, and the comparison section says “No publicly available API.”
Reasoning tiers and parameters: High, Medium, and Low. Temperature, sample count, prompts, and other run parameters are not specified.
The page uses Artificial Analysis Intelligence Index v5.0.1. The index consists of nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1. A higher index score is better.
| Reasoning tier | Intelligence Index | Time per index task (minutes, lower is better) | Output speed (tokens/s) | Weighted cost per index task (USD, lower is better) |
|---|---|---|---|---|
| High | 60 | 32.1 | 64 | $5.48 |
| Medium | 58 | 28.0 | 61 | $4.68 |
| Low | 54 | 25.9 | 56 | $5.22 |
The page summary also gives the High tier's time to first answer token as 73.49 seconds. High ranks 4th out of 81 models for Intelligence on this page. The page describes this tier as leading on the intelligence index, taking relatively long to complete tasks, and producing more output. Corresponding latency and ranking data for Medium and Low are not available in the currently visible data.
Total tokens in the index evaluation: The High-tier page summary shows 7.62B input tokens, an output-to-input ratio of 2.686%, and 205M output tokens for the index evaluation. These are usage figures for the index evaluation, not token counts for an individual task.
List price per million tokens: The High-tier page shows $2.00 for input and $10.00 for output, and lists a 95% cache discount. The page says these are prices per 1M input/output tokens.
Cost definitions: Cost per Intelligence Index task is a weighted average. The page also has a chart for the cost of running the full Intelligence Index, but that value is not shown in the current text view. List price per million tokens, weighted cost per individual task, and total cost for the full index are different metrics.
Context window: 1M tokens. The page also shows a Price field of $1.47 in the High, Medium, and Low table, but does not specify the unit next to that field. It is therefore not treated as interchangeable with the input/output token prices above.
This snapshot shows that higher tiers score better on the index, but also take longer. The per-task costs shown on the page do not increase monotonically with tier: Low costs $5.22, less than High at $5.48, but more than Medium at $4.68. Read each cost according to the page's own definition, and do not use these figures to infer the total cost of running the full index.
The page does not disclose the sample count, per-question outputs, specific prompts, calling parameters, or full execution environment for each tier of this model. This note records dynamic page data visible on 2026-11-16. The values may change on later visits, and the page alone is not sufficient to reproduce the experiment.
Open the original model page on Artificial Analysis and confirm that the model name and page path are gemini-4-carbon. Then review Intelligence Index v5.0.1, the tier comparison, task duration, and cost charts. When checking the results, record the access date and the index version shown on the page. The undisclosed sample count, prompts, and calling parameters cannot be recovered from this page.
Gemini 4 Carbon