Google DeepMindLanguage modelRestricted early access

Gemini 4 Carbon

Gemini 4 Carbon is among the strongest models: Rank 1 to 5 of 95 in the overall index.

IntelligenceRank 1 to 5 of 95

96Overall indexhigher is better

SpeedRank 48 of 98

100Tokens per secondhigher is better

Priceno API price

n/aper 1M tokenslower is better

ContextRank 30 of 99

1Mtokens at once≈ 1,500 A4 pages

Summary

At 100 tokens per second it writes faster than a typical model (97). Estimated without own results in: Mathematics and General.

Google announced Gemini 4 Carbon for longer software-engineering tasks and cybersecurity defense. It follows Gemini 4 Argon within the Gemini 4 family. The published provider evaluations include finding, validating, and fixing security vulnerabilities.

Early access starts with selected cybersecurity partners through the Fairwind Program. General access for developers, businesses, and consumers has been announced. Its introductory prices are documented separately from already available public API tariffs.

Gemini 4 Carbon benchmarks

Overall index (Gradually Meta-Score v3)

20 of 95 models

Claude Opus 5.596.4

Gemini 4 Carbon95.7

Claude Fable 5.194.9

Claude Sonnet 5.593.5

Gemini 4 Argon92.8

GPT-6 Astra91.9

GPT-6.1 Sol89.8

Muse Wonder Full89.5

Claude Opus 587.6

Muse Wonder82.1

Kimi K3.1 Pro79.9

Claude Fable 579.6

GPT-6 Sol73.4

GPT-5.6 Sol71.4

Gemini 3.7 Flash69.6

Muse Spark 1.369.3

GPT-6.1 Terra67.0

Gemini 3.8 Flash66.7

DeepSeek V4.1 Flash 101763.7

Grok 4.758.8

Gemini 4 Carbon

DeepSeek V4.1 Flash 1017

Grok 4.7

Speed

20 of 98 models

Mercury 2.5596tokens/s

Gemini 3.5 Flash-Lite454tokens/s

Gemini 3.7 Flash306tokens/s

Nemotron 3.5 Lightning298tokens/s

o3-mini281tokens/s

GPT-4.1 nano276tokens/s

Ling 3.0 Flash243tokens/s

Claude Haiku 5.5241tokens/s

GPT-5.4 mini230tokens/s

GPT OSS 120B225tokens/s

Inkling218tokens/s

Gemini 3 Flash Preview206tokens/s

Inkling Small204tokens/s

Qwen 3.7 Max203tokens/s

Gemini 3.6 Flash202tokens/s

Muse Spark 1.3198tokens/s

Muse Wonder197tokens/s

Gemini 3.5 Flash195tokens/s

Step 5 Preview182tokens/s

Gemini 4 Carbon100tokens/s

Mercury 2.5

Gemini 3.5 Flash-Lite

Gemini 3.7 Flash

Nemotron 3.5 Lightning

Gemini 3 Flash Preview

Gemini 4 Carbon

Output tokens per second, measured by Artificial Analysis

Gemini 4 Carbon by task area

0 to 100 · higher is better

Knowledgetoo little dataabove 99

ReasoningRank 1 to 6 of 8596

LegalRank 1 to 4 of 6694

FinanceRank 1 to 3 of 9694

CodingRank 3 to 7 of 8792

AgentsRank 1 to 6 of 8088

Light bar: too few results for a rank yet. No tests yet: Mathematics, General, Multimodal, Health.

Vibe Code Bench

Share of apps the model builds correctly from a description, in percent · higher is better

20 of 52 models

Claude Sonnet 5.592.4 %

Gemini 4 Argon91.9 %

Claude Haiku 5.590.4 %

Claude Fable 590.4 %

Claude Opus 5.590.3 %

Claude Fable 5.190.3 %

GPT-6 Astra89.6 %

Gemini 4 Carbon89.1 %

GPT-6.1 Sol88.9 %

Claude Opus 588.4 %

GPT-6 Sol87.8 %

DeepSeek V4.1 Flash 101787.7 %

Grok 4.786.2 %

Muse Wonder Full86.0 %

Muse Spark 1.385.9 %

Kimi K385.0 %

Kimi K3.1 Pro84.9 %

DeepSeek-V4.1-Flash84.7 %

GPT-6.1 Terra84.0 %

Muse Wonder83.1 %

Gemini 4 Carbon

DeepSeek V4.1 Flash 1017

as of October 28, 2026 · Source: Vals AI

Show all 502 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

AutomationBench-AA: Overall (Show measurement, test conditions, and source)

Source value: 77.81660572624064
Score: 77.81660572624064
Metric: Objectives completed without guardrail violations
Unit: %
Benchmark version: AA private 657-task split of AutomationBench v1.0.6
Category: agentic
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-automationbench-aa:AA private 657-task split of AutomationBench v1.0.6:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 77.8%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

CritPt v2 (Hard Subset): Overall (Show measurement, test conditions, and source)

Source value: 63
Score: 63
Metric: pass@1 accuracy
Unit: %
Benchmark version: v2 hard subset
Category: reasoning
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-critpt-v2-hard-subset:v2 hard subset:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 63.0%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high
Context: This is the 40-question hard subset of CritPt v2, not the full 70-question CritPt v2 run.

Source Artificial Analysis

CritPt v2: Overall (Show measurement, test conditions, and source)

Source value: 36
Score: 36
Metric: pass@1 accuracy
Unit: %
Benchmark version: v2
Category: reasoning
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-critpt-v2:v2:standalone:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 36.0%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

FrontierCode: Overall (Show measurement, test conditions, and source)

Source value: 40.560606060606105
Score: 40.560606060606105
Metric: Mean score over three attempts per task
Unit: %
Benchmark version: 2.0
Category: coding
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-frontiercode-2-0:2.0:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 40.6%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

FrontierCode 1.1 Main: Overall (Show measurement, test conditions, and source)

Source value: 48.5231333333333
Score: 48.5231333333333
Metric: Mean score over three attempts per task
Unit: %
Benchmark version: 1.1 Main
Category: coding
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-frontiercode:1.1 Main:standalone:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The primary 100-task FrontierCode 1.1 set is run three times per task, with equal task weights. Failing any blocking check or being flagged by the trajectory scanner scores zero; runs passing every blocking check score 70 to 100 percent.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Harness: Artificial Analysis Stirrup. Agent: Stirrup coding agent. Tool policy: 250-turn cap; repository worktree graded as left by the agent. Attempts: Mean score, 3 attempts per task. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 48.5%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The primary 100-task FrontierCode 1.1 set is run three times per task, with equal task weights. Failing any blocking check or being flagged by the trajectory scanner scores zero; runs passing every blocking check score 70 to 100 percent.. Harness: Artificial Analysis Stirrup. Agent: Stirrup coding agent. Tool policy: 250-turn cap; repository worktree graded as left by the agent. Attempts: Mean score, 3 attempts per task. Reasoning: reasoning enabled. Reasoning effort: high
Context: FrontierCode 1.1 Main is a private 100-task subset of 150 tasks. These are AA mean rubric scores, not a binary pass rate.

Source Artificial Analysis

GDP.pdf: Overall (Show measurement, test conditions, and source)

Source value: 28.4
Score: 28.4
Metric: All-pass rate
Unit: %
Benchmark version: unspecified
Category: reasoning
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-gdp-pdf:unspecified:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 28.4%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

HLE-Diamond: Overall (Show measurement, test conditions, and source)

Source value: 69.1
Score: 69.1
Metric: pass@1 accuracy
Unit: %
Benchmark version: Diamond
Category: general
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-hle-diamond:Diamond:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 69.1%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

AA-Omniscience Accuracy: Overall (Show measurement, test conditions, and source)

Source value: 58.266666666666666
Score: 58.266666666666666
Metric: AA-Omniscience Accuracy
Unit: %
Benchmark version: AA Omniscience diagnostic
Category: knowledge
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-omniscience-accuracy:AA Omniscience diagnostic:standalone:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Proportion of all questions answered correctly, regardless of whether the model answers or abstains on other questions.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Attempts: AA-reported aggregate; diagnostic-specific run count is not stated on the leaderboard. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 58.3%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Proportion of all questions answered correctly, regardless of whether the model answers or abstains on other questions.. Attempts: AA-reported aggregate; diagnostic-specific run count is not stated on the leaderboard. Reasoning: reasoning enabled. Reasoning effort: high
Context: This is the separately published accuracy diagnostic, not the AA-Omniscience Index. AA identifies higher values as better.

Source Artificial Analysis

AA-Omniscience Hallucination Rate: Overall (Show measurement, test conditions, and source)

Source value: 11.421725239616613
Score: 11.421725239616613
Metric: AA-Omniscience Hallucination Rate
Unit: %
Benchmark version: AA Omniscience diagnostic
Category: knowledge
Task: overall
Task label: Overall
Direction: Lower is better
Comparison cohort: artificial-analysis-omniscience-hallucination-rate:AA Omniscience diagnostic:standalone:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Proportion of incorrect answers among all non-correct responses: incorrect / (incorrect + partial answers + not attempted).. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Attempts: AA-reported aggregate; diagnostic-specific run count is not stated on the leaderboard. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 11.4%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Proportion of incorrect answers among all non-correct responses: incorrect / (incorrect + partial answers + not attempted).. Attempts: AA-reported aggregate; diagnostic-specific run count is not stated on the leaderboard. Reasoning: reasoning enabled. Reasoning effort: high
Context: This is the separately published hallucination-rate diagnostic, not the AA-Omniscience Index. AA identifies lower values as better.

Source Artificial Analysis

Terminal-Bench: Overall (Show measurement, test conditions, and source)

Source value: 65.1515151515151
Score: 65.1515151515151
Metric: Task success
Unit: %
Benchmark version: 4.0
Category: coding
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-terminal-bench-4:4.0:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 65.2%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

Terminal-Bench-Science: Overall (Show measurement, test conditions, and source)

Source value: 64.7619047619048
Score: 64.7619047619048
Metric: pass@1 task success rate
Unit: %
Benchmark version: 0.1.0
Category: reasoning
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: artificial-analysis-terminal-bench-science:0.1.0:v5.0:reasoning:effort=high:suffix=High
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-29
Methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Source model ID: artificial-analysis/909ed0ba-bcc5-4f38-8162-08a4e187a772/gemini-4-carbon/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Model snapshot: Gemini 4 Carbon (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 64.8%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-10-29-5a1b1cf41cd1. Source payload hash: 5a1b1cf41cd14d4fa7e1b5c737732505ff7f014395c8a19ff9385c13df09b106. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-10-29
Comparison methodology: Exact AA configuration: Gemini 4 Carbon (High) (release gemini-4-carbon, row gemini-4-carbon, UUID 909ed0ba-bcc5-4f38-8162-08a4e187a772, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 0. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap.. Reasoning: reasoning enabled. Reasoning effort: high

Source Artificial Analysis

EMB v1.1: Comps (Show measurement, test conditions, and source)

Source value: 83.067
Score: 83.067
Metric: accuracy
Unit: %
Benchmark version: 1.1
Category: finance
Task: comps
Task label: Comps
Direction: Higher is better
Comparison cohort: vals-emb-v1-1:1.1:comps
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-10-28
Methodology: Source model ID: google/gemini-4-carbon. Source model name: Gemini 4 Carbon. Provider reported by source: Google. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-10-27. Reasoning effort: high. Maximum output tokens: 1048576. Latency in seconds: 1643.738. Standard error: 3.981. Cost per task in USD: 5.799873. Source ID: vals-ai. Evidence capture ID: vals-emb-v1-1-2026-10-28-35dd0a8f6283. Source payload hash: 35dd0a8f6283a30582199f8ff0ca9f384f930274f8e480cb30f40eef2d6802fe. Parser version: vals-astro-v4. Imported at: 2026-10-28
Comparison methodology: Reasoning effort: high. Maximum output tokens: 1048576

Source Vals AI

Page 1 of 42

Head-to-head comparisons

vs.Gemini 4 Carbon vs. Claude Opus 5.5Level on performance.Gemini 4 CarbonClaude Opus 5.5Performance9696Pricen/a$8Speed10098 vs.Gemini 4 Carbon vs. Claude Sonnet 5.5Level on performance.Gemini 4 CarbonClaude Sonnet 5.5Performance9693Pricen/a$4Speed100138 vs.Gemini 4 Carbon vs. Claude Haiku 5.5Gemini 4 Carbon is stronger.Gemini 4 CarbonClaude Haiku 5.5Performance9657Pricen/a$0.20Speed100241 vs.Gemini 4 Carbon vs. Claude Fable 5.1Level on performance.Gemini 4 CarbonClaude Fable 5.1Performance9695Pricen/a$20Speed10065 vs.Gemini 4 Carbon vs. Claude Opus 5Gemini 4 Carbon is stronger.Gemini 4 CarbonClaude Opus 5Performance9688Pricen/a$10Speed10053 vs.Gemini 4 Carbon vs. Claude Fable 5Gemini 4 Carbon is stronger.Gemini 4 CarbonClaude Fable 5Performance9680Pricen/a$20Speed10070

Show all comparisons (57)

Gemini 4 Carbon vs. Claude Sonnet 5

Gemini 4 Carbon vs. Claude Opus 4.8

Gemini 4 Carbon vs. Claude Opus 4.7

Gemini 4 Carbon vs. Claude Opus 4.6

Gemini 4 Carbon vs. Claude Opus 4.5

Gemini 4 Carbon vs. Claude Sonnet 4.6

Gemini 4 Carbon vs. Claude Sonnet 4.5

Gemini 4 Carbon vs. Claude Haiku 4.5

Gemini 4 Carbon vs. GPT-6.1 Sol

Gemini 4 Carbon vs. GPT-6.1 Terra

Gemini 4 Carbon vs. GPT-6 Astra

Gemini 4 Carbon vs. GPT-6 Sol

Gemini 4 Carbon vs. GPT-6 Luna

Gemini 4 Carbon vs. GPT-5.6 Sol

Gemini 4 Carbon vs. GPT-5.6 Terra

Gemini 4 Carbon vs. GPT-5.6 Luna

Gemini 4 Carbon vs. GPT-5.5

Gemini 4 Carbon vs. GPT-5.4

Gemini 4 Carbon vs. GPT-5.4 mini

Gemini 4 Carbon vs. Gemini 4 Argon

Gemini 4 Carbon vs. Gemini 3.8 Flash

Gemini 4 Carbon vs. Gemini 3.7 Flash

Gemini 4 Carbon vs. Gemini 3.6 Flash

Gemini 4 Carbon vs. Gemini 3.5 Flash

Gemini 4 Carbon vs. Gemini 3.1 Pro Preview

Gemini 4 Carbon vs. Gemini 3.5 Flash-Lite

Gemini 4 Carbon vs. Muse Wonder

Gemini 4 Carbon vs. Muse Wonder Full

Gemini 4 Carbon vs. Muse Spark 1.3

Gemini 4 Carbon vs. Grok 4.7

Gemini 4 Carbon vs. Grok 4.6

Gemini 4 Carbon vs. Grok 4.5

Gemini 4 Carbon vs. Grok 4.3

Gemini 4 Carbon vs. DeepSeek-V4-Pro

Gemini 4 Carbon vs. DeepSeek-V4-Flash

Gemini 4 Carbon vs. DeepSeek-V4.1-Flash

Gemini 4 Carbon vs. DeepSeek V4.1 Flash 1017

Gemini 4 Carbon vs. Kimi K3.1 Pro

Gemini 4 Carbon vs. Kimi K3

Gemini 4 Carbon vs. Kimi K2.6

Gemini 4 Carbon vs. GLM-5.3-Flash

Gemini 4 Carbon vs. GLM-5.3

Gemini 4 Carbon vs. GLM-5.2

Gemini 4 Carbon vs. GLM-5.1

Gemini 4 Carbon vs. Qwen 3.8 Max 0902

Gemini 4 Carbon vs. Qwen 3.8 27B

Gemini 4 Carbon vs. Mistral Large 4

Gemini 4 Carbon vs. Mistral Large 3

Gemini 4 Carbon vs. Mistral Medium 3.5

Gemini 4 Carbon vs. MiMo-V2.5

Gemini 4 Carbon vs. MiMo-V2.5-Pro

All specifications

Input and output

Context window: 1,000,000 tokens; Source: Context window
Inputs reported by Artificial Analysis: Text, image, audio, and video; Source: Inputs reported by Artificial Analysis
Output reported by Artificial Analysis: Text; Source: Output reported by Artificial Analysis

Licence and access

Weights: Proprietary; Source: Weights

Further details

Announcement: October 28, 2026; Source: Announcement
Availability: Initially for selected cybersecurity partners through the Fairwind Program. General access for developers, businesses, and consumers has been announced, without a public launch date.; Source: Availability
Context-window source: Artificial Analysis reports 1,000,000 tokens. Google does not publish an exact context limit in its announcement.; Source: Context-window source
AA rank reported by the developer: Rank #1 on Artificial Analysis Intelligence Index v5.0 (score 59). (Rank reported at Google’s announcement on October 28, 2026; not a continuously updated ranking.); Source: AA rank reported by the developer
WebDev Arena rank reported by the developer: Rank #4 on WebDev Arena. (Rank reported at Google’s announcement on October 28, 2026; not a continuously updated ranking.); Source: WebDev Arena rank reported by the developer
WebDev Arena change reported by the developer: +89 Elo points reported. (This is a reported score change, not an absolute score; no starting rating is supplied in the announcement.); Source: WebDev Arena change reported by the developer
Secure code generation reported by the developer: More than 2x uplift in secure-code-generation success rate versus Argon. (Relative internal comparison only; neither absolute rate is published here, and this is not a percentage-point change.); Source: Secure code generation reported by the developer
Chrome patches reported by the developer: Chrome team reports more than 10x better high-quality patch generation than another lab’s large frontier model. (The comparator model, metric definition, and absolute scores are not specified; retain this as a vendor-reported relative claim.); Source: Chrome patches reported by the developer
Misalignment alerts reported by the developer: High-severity misalignment alerts dropped by almost 80% in Google’s live coding environments. (Relative approximate reduction; absolute alert rates, comparison baseline, and period are not stated. Not percentage points.); Source: Misalignment alerts reported by the developer
Unsupported claims reported by the developer: Internal testing says Carbon cuts unsupported claims by approximately 2x versus Argon. (Relative internal comparison only; absolute rates and evaluation details are not supplied.); Source: Unsupported claims reported by the developer
Overreach reported by the developer: Reported 26% reduction in overreach. (Relative reduction as stated, not percentage points; the announcement does not state an absolute score or comparator in this sentence.); Source: Overreach reported by the developer
Under-service reported by the developer: Reported 29% reduction in under-service. (Relative reduction as stated, not percentage points; the announcement does not state an absolute score or comparator in this sentence.); Source: Under-service reported by the developer
Agent Arena model sessions: 26,511 (Published total model sessions. Arena does not publish separate sample sizes for the individual signals.) (retrieved 2026-10-29); Source: Agent Arena model sessions

Prices

Announced introductory pricing: $2 input and $10 output per 1M tokens. Cached input receives a 95% discount, giving $0.10 per 1M tokens. This announcement does not establish an already available public API.Source
Announced pricing after the introductory period: Afterward, $4 input and $20 output per 1M tokens. Google has not specified when the introductory period ends.Source
Artificial Analysis pricing assumptions: Artificial Analysis calculates test costs using $4 input, $20 output, and $0.20 cached input per 1M tokens. These are the evaluator’s pricing assumptions, not confirmation of a publicly available API tariff.Source

Frequently asked questions about Gemini 4 Carbon

How good is Gemini 4 Carbon in benchmarks?

Very good: Rank 1 to 5 of 95 in the overall index.

Overall index (Gradually Meta-Score v3) from 19 benchmark results · data through Oct 29, 2026

IntelligenceOverall index, 0 to 100 · higher is better

Claude Opus 5.596

Gemini 4 CarbonRank 1 to 596

Claude Fable 5.195

Claude Sonnet 5.593

Gemini 4 Argon93

GPT-6 Astra92

GPT-6.1 Sol90

What is the Overall index (Gradually Meta-Score v3)?

A value from 0 to 100 that combines many independent benchmarks into one number. Gemini 4 Carbon scores 96 out of 100, which is rank 1 to 5 of 95.

Own analysis by gradually.ai, method benchmark-meta-score-v3 · for Gemini 4 Carbon from 19 test series

Which data: Only current, reviewed results from independent evaluators such as Artificial Analysis, Vals AI, LMArena, LiveBench and Epoch AI. Vendor claims and historical or disputed values do not count.
One value per test: For each source and test series the latest retrieval with one fixed setup counts. When a source tests several reasoning levels, the highest documented one applies. A series needs at least 5 models.
Making tests comparable: Percentages are converted so that gaps near 0 and near 100 % count fairly, then every series is put on the same scale. A hard test therefore does not weigh more than an easy one.
Shared strength: A statistical model estimates, from all series at once, one shared capability per model and, per area, how far a model deviates from it. Missing tests are never counted as zero.
Weighting: Core areas (Coding, Reasoning, Mathematics, Knowledge, Agents, General and Multimodal) count 75 % together, the domain areas Finance, Legal and Health 25 %.
Scale from 0 to 100: The value is measured against a fixed reference group of 29 models: 50 is its middle, above 99 and below 1 are the outer ends. New models therefore do not shift older values.
Place instead of decimals: The score is recomputed 200 times with randomly redrawn test families. The place gives the range the model almost always lands in, such as "rank 4 to 8". Two models level within that range are not reliably different.
Minimum evidence: A value needs at least 8 test series from 2 sources and 3 areas, a place only with enough precision. Missing core areas are flagged as "estimated".
What it is not: Not a test run by gradually.ai and no verdict on every single task. For a specific task the area values and the individual benchmarks matter more.

How large is the context window of Gemini 4 Carbon?

1 M tokens. That is how much text Gemini 4 Carbon reads at once, about 1,500 A4 pages.

Page count as a rule of thumb from Artificial Analysis

Context windowTokens the model reads at once · higher is better

GPT-6 Astra1.05M

GPT-6.1 Sol1.05M

Gemini 4 Argon1.05M

Gemini 4 CarbonRank 30 of 991M

Claude Opus 5.51M

Claude Fable 5.11M

Claude Sonnet 5.51M

How fast is Gemini 4 Carbon?

100 tokens per second. That is faster than a typical model at 97.

Measured by Artificial Analysis on Oct 28, 2026

SpeedTokens per second · higher is better

Claude Sonnet 5.5138

Gemini 4 CarbonRank 48 of 98100

Claude Opus 5.598

Gemini 4 Argon82

GPT-6.1 Sol79

GPT-6 Astra66

Claude Fable 5.165

Can I run Gemini 4 Carbon locally?

No. Google DeepMind does not publish the weights.

Stated by Google DeepMind

Open weights: no
License: n/a
Parameters: n/a
active per token: n/a

Where can I use Gemini 4 Carbon?

Initially for selected cybersecurity partners through the Fairwind Program. General access for developers, businesses, and consumers has been announced, without a public launch date.

Stated by Google DeepMind

Availability

Initially for selected cybersecurity partners through the Fairwind Program. General access for developers, businesses, and consumers has been announced, without a public launch date.

Is Gemini 4 Carbon better than Gemini 4 Argon?

Not clearly in the overall index: both are level. It leads in Knowledge and Agents.

Gemini 4 Argon as earlier generation according to Google: Introducing Gemini 4 Carbon, checked Oct 28, 2026

IntelligenceOverall index, 0 to 100 · higher is better

Gemini 4 Carbon

Rank 1 to 5

Gemini 4 Argon

Rank 2 to 7

SpeedTokens per second · higher is better

100

Gemini 4 Carbon

fairly fast

Gemini 4 Argon

fairly slow

PriceDollars per 1M tokens · lower is better

n/a

Gemini 4 Carbon

Gemini 4 Argon

fairly expensive

Area4 Carbon4 Argon0 to 100, higher is betterResult

Codinglevel9290

Reasoninglevel9690

Mathematicsonly 4 Argon has a scorenot enough data90

Knowledge4 Carbon aheadabove 9997

Agents4 Carbon ahead8880

Generalonly 4 Argon has a scorenot enough data92

Financelevel9492

Legallevel9491

Healthonly 4 Argon has a scorenot enough data93

Sources and data date

Show all 61 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

Evaluator	Test series	Composite indices	Recorded active measurements
Artificial Analysis	10	1	100
Vals AI	9	1	335
LMArena	0	0	42

Without vendor claims and historical or disputed results; those stay labelled in the single values. 127 measurements are not yet assigned to a test series.

Type	Evidence and data date
Speed	Google, High effort. Output speed after the first streaming chunk, not total response time. Artificial Analysis early-access measurement through Google, with 10,000 input tokens and one concurrent request. Carbon has restricted access; the measurement does not establish public API availability. Retrieved October 28, 2026. Source: Artificial Analysis · Context window data (retrieved October 28, 2026)Speed measurement methodology
Input	Documented modalities. Source: Context-window data · Context window data (retrieved October 28, 2026)
Output	Documented modalities. Source: Context-window data · Context window data (retrieved October 28, 2026)
Research status	Research date October 28, 2026. 98 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	Carbon-specific Safety and Security report Note The announcement links the Gemini 4 Argon Safety & Security Report, not a Carbon report. No Carbon-specific report results were used.
Not published	Per-result sample counts and confidence intervals Note The announcement table, charts, and text do not publish sample counts or numeric uncertainty intervals for the Carbon observations.
Not published	Absolute scores and comparator details for internal relative claims Note The announcement gives relative or approximate internal changes for secure-code generation, Chrome patching, misalignment alerts, unsupported claims, overreach, and under-service, but not the corresponding absolute scores; only the explicitly named Argon comparisons are labeled as such.
Not published	Detailed run configurations for table results without row-specific labels Note The announcement provides the table values and a few row sublabels, but does not state prompts, sampling settings, or other detailed run configurations for most general benchmarks.
Not found in the inspected sources	Vals Code Migration v1.1 standalone exact result Note The captured standalone Code Migration v1.1 JSON lists neither google/gemini-4-carbon nor a Carbon score. Its unique Vals Index component is imported by the parent into the central database; it is not represented as a standalone leaderboard result.
Unresolved	Vals accuracy-score uncertainty metadata Note The parent imports the exact Vals accuracy rows into the central benchmark database. Vals publishes numeric stderr but the captured pages and payloads do not define its statistical meaning or confidence level, and no accuracy-score denominator is published. token_tasks_measured and turn_metrics.tasks_measured are coverage for token/turn diagnostics, not accuracy-score sample counts.
Not found in the inspected sources	LiveBench exact Carbon result Note No exact google/gemini-4-carbon model row is present in the saved LiveBench 2.0 2026-10-25 homepage/table/categories/breakdown/notes captures. This is limited to the saved release; later LiveBench releases were not refreshed.
Not found in the inspected sources	Scale Labs exact Carbon result Note No exact Gemini 4 Carbon row was present in the 20 saved Scale Labs leaderboard pages checked on 2026-10-28. Scope is those captured evaluator pages only.
Not found in the inspected sources	DesignArena exact Carbon result Note No exact Gemini 4 Carbon row was present in the saved DesignArena leaderboard page checked on 2026-10-28.
Not found in the inspected sources	DualEntry exact Carbon result Note No exact Gemini 4 Carbon row was present in the saved DualEntry accounting benchmark page checked on 2026-10-28.
Not found in the inspected sources	Plotly DataAnalyticsBench exact Carbon result Note No exact Gemini 4 Carbon row was present in the saved Plotly DataAnalyticsBench article checked on 2026-10-28.
Not found in the inspected sources	Arena WebDev Fullstack exact row Note The official Fullstack page reports 88 models and 115,229 votes. Its complete captured page text was searched on 2026-10-29 and contains no Gemini 4 Carbon row. The current page snapshot is dated 2026-10-27.
Unresolved	Remaining Arena Text category pages Note Exact rows were captured for Math, Instruction Following, Multi-Turn, Creative Writing, Coding, Hard Prompts, Hard Prompts (English), Longer Query, Non-English, Chinese, Expert, seven accessible occupational categories, Russian, and the Excluding Ties view. German and Japanese full pages were checked and had no Carbon row. Spanish, French, Korean, Arabic, Hindi, Portuguese, Italian, Thai, Indonesian, and Occupational Software & IT Services pages were inaccessible to the official-page reader. Arena labels the Text category directory as 29 categories, but its collapsed navigation does not expose every remaining label in the captured page text. Unresolved pages are not treated as negative results.
Unresolved	Arena leaderboards outside Agent, Text and Code Arena WebDev Note The official changelog says Gemini 4 Carbon was added to Agent, Text and Code Arena WebDev on 2026-10-28. Arena Vision, Image-to-WebDev, Document, Search, Image, and Video leaderboards were not individually captured for exact Carbon rows in this task; no absence claim is made for those leaderboards.
Unresolved	Arena WebDev page-date discrepancy Note Arena’s current WebDev pages display Oct 27, 2026, while the official changelog says Gemini 4 Carbon was added to Code Arena: WebDev on Oct 28, 2026. Both dates are preserved as shown; no explanation was published in the checked pages.
Not published	Vals latency and cost sample denominators / confidence intervals Note The 214 retained Vals operational rows comprise 107 latency and 107 task-cost observations. Their captured payloads do not publish latency sample counts or confidence intervals, nor the exact cost-sampling denominator or a cost confidence interval. Token/turn coverage fields are retained in configuration but are not treated as denominators for these metrics.
Additional source	Model metadata (retrieved October 28, 2026) · Google: Introducing Gemini 4 Carbon · Editorial description reviewed October 28, 2026
Additional source	Artificial Analysis (retrieved October 29, 2026)1 2 3 4 5 6 7 8 9 10 11
Additional source	Vals AI (retrieved October 28, 2026)1 2 3 4 5 6 7 8
Additional source	Google Gemini 4 Carbon announcement, published 2026-10-28 (retrieved October 28, 2026)1 2 3 4 5
Additional source	Arena.ai (retrieved October 29, 2026)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
Additional source	Arena.ai Agent Arena (retrieved October 29, 2026)