modelbench日本語

Gemini 4 Carbon

GoogleLimited availability

Input price
$2.00
Per 1 million tokens (USD)
Output price
$10.00
Per 1 million tokens (USD)
Effort levels
3total
Default: Med
Benchmark results
36total
Benchmark types: 29

Model information

API model ID
gemini-4-carbon
release date
2026-10-28
effort
Low / Med (default) / High
Effort source
Based on thinking level in Gemini API thinking documentation (aistudio.google.com/docs/thinking).
Claude Code
unavailable
Codex
unavailable
Availability notes
Not generally available. At announcement, only some cyber defenders and trusted partners in the Fairwind Program (ID listed in the pricing table). The official model list for Claude Code / Codex (checked 2026-11-12) contains only Anthropic / OpenAI models, respectively, so both are false.
Availability source
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/

Price (USD per 1 million tokens)

Prices do not vary by effort; token consumption does.

input$2.00
output$10.00
cache read$0.10
cache write—

Introductory price. Revised to input $4.00 / output $20.00 / cache $0.20 from 2026-12-01 (arrow notation in the pricing table). Cache is 95% off

Gemini 4 Carbon gemini-4-carbon Input $2.00 → $4.00 from Dec 1, 2026 / Output $10.00 → $20.00 / Context caching $0.10 → $0.20

Source: https://aistudio.google.com/docs/pricing

Benchmark results

BenchmarkeffortScoreSource cost (metric)measured/reported bysourceConditions and extraction method
Terminal-Bench 4.0HighHigh65.2%—Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: a measured value from Artificial Analysis (labeled Terminal-Bench 4.0 (AA))transcribed from table
Source excerptTerminal-Bench 4.0: Gemini 4 Carbon 65.2
Terminal-Bench-Science 0.1not statednot stated64.8%—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt64.8%Location Table, Terminal-Bench-Science 0.1 row: 64.8
Terminal-Bench-Science 0.1HighHigh64.8%—Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: a measured value from Artificial Analysistranscribed from table
Source excerptTerminal-Bench Science 0.1: Gemini 4 Carbon 64.8
FrontierCode 2.0LowLow32.4%$5.49Cost per tasknot statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. A point in the figure "FrontierCode 2.0 / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Low" was mapped to low in the Gemini API thinking leveltranscribed from text
Source excerptFrontierCode 2.0 — Accuracy vs. cost: "Gemini 4 Carbon · Low: 32.4% $5.49"
FrontierCode 2.0Med (default)Med41.1%$5.16Cost per tasknot statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. A point in the figure "FrontierCode 2.0 / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "Med" was mapped to medium in the Gemini API thinking leveltranscribed from text
Source excerptFrontierCode 2.0 — Accuracy vs. cost: "Gemini 4 Carbon · Med: 41.1% $5.16"
FrontierCode 2.0HighHigh38.4%$5.83Cost per tasknot statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. A point in the figure "FrontierCode 2.0 / Accuracy vs. cost" on the Anthropic announcement page. values were taken from each point's aria-label (DOM); the page shows them only in the figure; value for a model from some other vendor (announced by Anthropic); the figure's effort label "High" was mapped to high in the Gemini API thinking leveltranscribed from text
Source excerptFrontierCode 2.0 — Accuracy vs. cost: "Gemini 4 Carbon · High: 38.4% $5.83"
FrontierSWE 2not statednot stated61.2%—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt61.2%Location Table, FrontierSWE v2 row: 61.2
FrontierSWE 2HighHigh61.2%—ProximalGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: values from the Proximal official leaderboardtranscribed from table
Source excerptFrontierSWE 2: Gemini 4 Carbon 61.2
SWE-sweepHighHigh9.6%—GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: self-measured for Carbon using the mini-swe-agent harnesstranscribed from table
Source excerptSWE-sweep: Gemini 4 Carbon 9.6
AA-Briefcase v1.1(Elo)not statednot stated1,611—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt1611Location Table, AA-Briefcase v1.1 (Elo) row: 1611
AA-Briefcase v1.1(rubric)not statednot stated61.8%—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt61.8%Location Table, AA-Briefcase v1.1 (rubric) row: 61.8
AutomationBenchnot statednot stated53.9%—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt53.9%Location Table, AutomationBench row: 53.9
AutomationBenchHighHigh53.9%—ZapierGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: Zapier official leaderboard (private set)transcribed from table
Source excerptAutomationBench: Gemini 4 Carbon 53.9
OSWorld 2.1 (offline set, partial)not statednot stated81.5%—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt81.5%Location Table, OSWorld 2.1 (offline set, partial) row: 81.5
OSWorld 2.1 (offline set, partial)HighHigh81.5%—GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: self-measured. 82-task offline subset, maximum of 3 runs, partial scoretranscribed from table
Source excerptOSWorld-2.1 (offline subset): Gemini 4 Carbon 81.5
Agents' Last Exam v1.1(rubric score)HighHigh77.4%—GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: self-measured. 122 tasks, ALE-Claw harness, 5-hour window, averaged rubric Scoretranscribed from table
Source excerptAgents' Last Exam v1.1: Gemini 4 Carbon 77.4
Chartography (without tools)not statednot stated72.5%—not statedIntroducing Claude Fable 5.2 and Claude Mythos 5.2official · Retrieved 2026-11-12Vendor-reported value. table note: Scores for Fable 5.2, Fable 5.1, and Opus 5.5 show results with safeguards on unless otherwise noted. All scores show each model’s highest-available effort unless otherwise noted. Values for other vendor models (as listed by Anthropic).transcribed from table
Source excerpt72.5%Location Table, Chartography (no tools) row: 72.5
Chartography (without tools)HighHigh72.5%—not statedGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: without tools, following the Surge official leaderboardtranscribed from table
Source excerptChartography: Gemini 4 Carbon 72.5
Artificial Analysis Intelligence Index v5.0HighHigh59—Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: values reproduced from Artificial Analysis (third-party)transcribed from table
Source excerptArtificial Analysis Intelligence Index v5.0: Gemini 4 Carbon 59
Vals Index v2.2HighHigh70.4%—Vals AIGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: values reproduced from Vals AItranscribed from table
Source excerptVals Index v2.2: Gemini 4 Carbon 70.4
Vending-Bench 2HighHigh16,484—Andon LabsGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: Andon Labs official leaderboard (final balance $)transcribed from table
Source excerptVending-Bench 2: Gemini 4 Carbon 16484
Blueprint-Bench 2HighHigh58.6%—Andon LabsGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: Andon Labs official leaderboard (similarity score)transcribed from table
Source excerptBlueprint-Bench 2: Gemini 4 Carbon 58.6
GraphWalks BFS (256k to 1M)HighHigh96.2%—GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: self-measured. 200 questions, BFS F1transcribed from table
Source excerptGraphWalks BFS (256k-1M): Gemini 4 Carbon 96.2
RiemannBenchHighHigh72.0%—Surge AIGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: Surge official leaderboardtranscribed from table
Source excerptRiemannBench: Gemini 4 Carbon 72.0
CWE-bench v1HighHigh71.0%—Collinear AI / Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: official leaderboard (independent evaluation by Collinear AI / Artificial Analysis), programmatic pass@1transcribed from table
Source excerptCWE-bench v1: Gemini 4 Carbon 71.0
ExploitGymHighHigh54.8%—GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: self-measured. Antigravity harness, single trial, no internet, default 6-hour timeout (performance via Fairwind Program)transcribed from table
Source excerptExploitGym: Gemini 4 Carbon 54.8
SEC-bench ProHighHigh88.2%—GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)official · Retrieved 2026-11-12Vendor-reported value. Google published table. methodology PDF: Gemini 4 Carbon/Argon used the highest Gemini API thinking setting, pass@1, and a single trial. Row-specific source: self-measured. 183 tasks (2026-05 JS engine split), mean Pass@1 over 3 trials, Antigravity harness (via Fairwind Program)transcribed from table
Source excerptSEC-bench Pro: Gemini 4 Carbon 88.2
GPQA DiamondHighHigh96.5%—Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0third-party · Retrieved 2026-11-12Run by Epoch AI (third-party measurement). Source label "Gemini 4 Carbon (high)". effort is indicated by the suffix _high in Epoch's Model version. Data retrieved 2026-11-12.transcribed from table
Source excerptgpqa_diamond.csv: Model version=gemini-4-carbon_high, Best score (across scorers)=0.9652777777777778, stderr=0.01207061012675132, Started at=2026-11-02T23:17:42.000Z
FrontierMath Tiers 1-3(v2)HighHigh90.5%—Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0third-party · Retrieved 2026-11-12Run by Epoch AI (third-party measurement). Source label "Gemini 4 Carbon (high)". effort is indicated by the suffix _high in Epoch's Model version. Data retrieved 2026-11-12.transcribed from table
Source excerptfrontiermath_tiers_1_3_v2.csv: Model version=gemini-4-carbon_high, Best score (across scorers)=0.9052631578947369, stderr=0.017377516319276178, Started at=2026-11-02T23:17:42.000Z
FrontierMath Tier 4(v2)HighHigh95.1%—Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0third-party · Retrieved 2026-11-12Run by Epoch AI (third-party measurement). Source label "Gemini 4 Carbon (high)". effort is indicated by the suffix _high in Epoch's Model version. Data retrieved 2026-11-12.transcribed from table
Source excerptfrontiermath_tier_4_v2.csv: Model version=gemini-4-carbon_high, Best score (across scorers)=0.9512195121951219, stderr=0.03405912205797303, Started at=2026-11-02T23:17:42.000Z
OTIS Mock AIME 2024-2025HighHigh100.0%—Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0third-party · Retrieved 2026-11-12Run by Epoch AI (third-party measurement). Source label "Gemini 4 Carbon (high)". effort is indicated by the suffix _high in Epoch's Model version. Data retrieved 2026-11-12.transcribed from table
Source excerptotis_mock_aime_2024_2025.csv: Model version=gemini-4-carbon_high, Best score (across scorers)=1.0, stderr=0.0, Started at=2026-11-02T23:17:42.000Z
SimpleQA VerifiedHighHigh77.7%—Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0third-party · Retrieved 2026-11-12Run by Epoch AI (third-party measurement). Source label "Gemini 4 Carbon (high)". effort is indicated by the suffix _high in Epoch's Model version. Data retrieved 2026-11-12.transcribed from table
Source excerptsimpleqa_verified.csv: Model version=gemini-4-carbon_high, Best score (across scorers)=0.777, stderr=0.013169830843425608, Started at=2026-11-02T23:17:42.000Z
Chess Puzzles(Epoch AI)HighHigh78.0%—Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0third-party · Retrieved 2026-11-12Run by Epoch AI (third-party measurement). Source label "Gemini 4 Carbon (high)". effort is indicated by the suffix _high in Epoch's Model version. Data retrieved 2026-11-12.transcribed from table
Source excerptchess_puzzles.csv: Model version=gemini-4-carbon_high, Best score (across scorers)=0.78, stderr=0.04163331998932266, Started at=2026-11-02T23:17:42.000Z
Epoch Capabilities Index(ECI)not statednot stated169.76—Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)third-party · Retrieved 2026-11-12A per-model metric that combines results across multiple benchmarks and multiple efforts (not per effort). Calculated by Epoch AI.transcribed from table
Source excerpteci_scores.csv: Model=Gemini 4 Carbon, eci=169.76, eci_ci_low=166.48, eci_ci_high=173.5
LMArena Text (style control, overall)HighHigh1,526—LMArenalmarena-ai/leaderboard-dataset text_style_control/latest-00000-of-00001.parquet (CC BY 4.0)third-party · Retrieved 2026-11-12Human head-to-head votes (Bradley-Terry). Source label "Gemini 4 Carbon (High)" (effort is in parentheses in the label). Release date 2026-11-09.transcribed from table
Source excerptmodel_name=Gemini 4 Carbon (High), rating=1525.95, rating_lower=1516.4, rating_upper=1535.6, vote_count=4282, rank=1, category=overall, leaderboard_publish_date=2026-11-09
LMArena Text Coding (without style control)HighHigh1,563—LMArenalmarena-ai/leaderboard-dataset text/latest-00000-of-00001.parquet (CC BY 4.0)third-party · Retrieved 2026-11-12Human head-to-head votes (Bradley-Terry). Source label "Gemini 4 Carbon (High)" (effort is in parentheses in the label). Release date 2026-11-09.transcribed from table
Source excerptmodel_name=Gemini 4 Carbon (High), rating=1563.41, rating_lower=1544.8, rating_upper=1582.0, vote_count=1072, rank=1, category=coding, leaderboard_publish_date=2026-11-09

Related rankings

References