gemini-4-carbon: capability breakdown
162.5 / 220
measured 6 of 22 rungs · effort row: high (the median of its published effort settings)
Provisional. Measured from Artificial Analysis's public page on 2026-11-11, High effort row (gemini-4-carbon), AA's default and the only full row. The personality capture ran at Google's provider-default reasoning. Fitted with the 2026-10-04 provisional method (rung difficulties and slopes of the 2026-09-29 ladder, without the full refit), which reproduces the published ladder to within 0.9 points on average across 148 scored models and tends to read 0.6 points low. 6 of 22 rungs are measured; the rest are fitted, so this score is less certain than most. The next full pipeline run replaces it.
This page shows where the number comes from. Each rung is one public benchmark, worth up to 10 points times the model's chance-corrected score on it; the score is the sum of all 22 rungs. How the capability ladder works explains the rungs, the fit and its limits.
| Rung (easiest → hardest) | Points / 10 | How | Source date |
|---|---|---|---|
| HumanEval | 9.97 | fitted | |
| MMLU | 9.83 | fitted | |
| MATH-500 | 9.94 | fitted | |
| MMLU-Pro | 9.63 | fitted | |
| AIME 2024 | 10.00 | fitted | |
| AIME 2025 | 9.93 | fitted | |
| LiveCodeBench | 9.59 | fitted | |
| GPQA Diamond | 9.46 | fitted | |
| IFBench | 7.52 | fitted | |
| τ²-Telecom | 9.65 | fitted | |
| AA-LCR | 7.93 | measured | 2026-11-11 |
| Terminal-Bench 2.1 | 8.69 | fitted | |
| SciCode | 5.97 | fitted | |
| Terminal-Bench Hard | 6.53 | fitted | |
| AutomationBench | 7.78 | measured | 2026-11-11 |
| AA-Briefcase (rubric pass) | 6.18 | measured | 2026-11-11 |
| HLE | 4.74 | fitted | |
| MLCR | 1.94 | measured | 2026-11-11 |
| τ³-Banking | 4.47 | fitted | |
| Omniscience acc. | 5.83 | measured | 2026-11-11 |
| CritPt | 4.06 | fitted | |
| GDP.pdf | 2.84 | measured | 2026-11-11 |
| Total | 162.48 | = 162.5 / 220 | |
Measured rows are the model's own published result on that benchmark. fitted rows are an item-response estimate: what a model with this ability would be expected to score on a benchmark it has no published result for (see the methodology).
Per-benchmark results are Artificial Analysis's (public model pages); fitted values are this site's estimates. Permission to republish requested 2026-09-05.