RankingEffort benchmarks

Gemini 4 Argon

Compare models

18 measured results · 1 effort settings · 18 benchmark/harness combinations

Model evidence and sources →

Change model

Model

Current configurationSources

Compare with

Choose a second model and its effort to see both profiles and benchmark differences below.

Compare withModel and configuration

Capability profile

One configuration, seven capabilities

Comparison score · 0–100

AGemini 4 Argon · High3 of 7 capabilities supported

Gaps are unknown, not zero. Hollow points are preliminary.

Agentic
A81.9Supported
Hard reasoning
A62.6Supported
Coding
A85.7Supported
Human pref
Unknown
Knowledge
A47.9Preliminary
Multimodal
Unknown
Long context
A51.1Preliminary

Each profile belongs to the named configuration. Capability scores use a shared panel for all axes, distinct from the overall ranking panel; they are not accuracy percentages or direct head-to-head wins.

Capability score evidence

Gemini 4 Argon · High: 79.7 overall comparison score. Meets the overall evidence requirements.

5 areas measured. The available task mix affects this estimate; missing areas remain unknown.

Which sources carry the weight?

Artificial Analysis · 70.1% direct comparison weight · independent of the model provider
Vals AI · 29.9% direct comparison weight · independent of the model provider

These shares describe the comparisons entering the fit, not fractions of the final score. Source balancing does not prove independence: overlapping sources may remain correlated.

How sensitive is this score?

The audit removes one source at a time and refits against the same overall reference panel. These published results use the matching dataset audited 2026-10-01.

Both ranks use the same surviving configurations for each removal. The cohort starts with 429 configurations supported in the original audit, including separate effort settings. These are diagnostic positions, not the default leaderboard ranks. Eligibility is not rechecked after removal.

What changes when a source is removed?
Removed source	Rank before → after	Rank movement	Cohort remaining / lost support	Score after (change)
Artificial Analysis	20 → 9	Up 11	193 remaining / 236 lost	90.0 (+10.3 points)
ARC Prize	20 → 15	Up 5	429 remaining / 0 lost	82.0 (+2.4 points)
Datacurve	20 → 21	Down 1	429 remaining / 0 lost	78.2 (-1.5 points)
LiveCodeBench	20 → 20	No movement	429 remaining / 0 lost	79.6 (-0.0 points)
Scale	20 → 19	Up 1	429 remaining / 0 lost	80.3 (+0.7 points)
Terminal-Bench	20 → 17	Up 3	429 remaining / 0 lost	80.4 (+0.7 points)
Cognition	20 → 17	Up 3	429 remaining / 0 lost	80.3 (+0.6 points)
LiveBench	20 → 18	Up 2	428 remaining / 1 lost	81.4 (+1.8 points)
SWE-rebench	20 → 20	No movement	429 remaining / 0 lost	79.4 (-0.3 points)
Epoch AI	20 → 20	No movement	412 remaining / 17 lost	80.6 (+1.0 points)
Vals AI	20 → 38	Down 18	428 remaining / 1 lost	65.5 (-14.1 points)
Provider reports · Anthropic	20 → 21	Down 1	429 remaining / 0 lost	79.6 (-0.1 points)
Provider reports · Moonshot AI	20 → 21	Down 1	429 remaining / 0 lost	79.4 (-0.3 points)
Provider reports · Meta	20 → 21	Down 1	429 remaining / 0 lost	79.3 (-0.4 points)
Provider reports · Mistral	20 → 20	No movement	429 remaining / 0 lost	79.7 (+0.0 points)
Provider reports · OpenAI	20 → 21	Down 1	429 remaining / 0 lost	79.4 (-0.3 points)
All provider reports	20 → 24	Down 4	429 remaining / 0 lost	78.7 (-1.0 points)

Lost support means a configuration no longer has a comparison path to the complete reference panel. Its rank is unavailable. A surviving path alone does not establish leaderboard eligibility.

Overlapping sources may remain correlated even with source balancing. Leave-one-source-out movement is a sensitivity diagnostic, not a spread of source-specific ranks or a confidence measure. Scoring methodology · Dated configuration audit

Every benchmark family

Removed evidence	Recomputed score	Change
arc-agi	82.0	+2.4
automationbench	78.1	-1.5
deepswe	78.2	-1.5
omniscience	81.0	+1.4
aa-lcr	81.0	+1.4
gdp-pdf	80.7	+1.0
mmlu-pro	80.5	+0.8
vals-finance-agent	78.9	-0.7
frontiercode	80.3	+0.6
arc-agi-3	79.1	-0.6
vals-tax-agent	79.1	-0.5
enigma-eval	80.1	+0.4
mmmu-pro	80.1	+0.4
terminal-bench	80.0	+0.4
ifbench	79.3	-0.3
critpt	80.0	+0.3
tau-bench	79.3	-0.3
vibe-code	79.4	-0.3
simpleqa	79.9	+0.3
swe-rebench	79.4	-0.3
epoch-game-puzzles	79.4	-0.3
hle	79.4	-0.3
swe-atlas-test-writing	79.4	-0.2
harvey	79.9	+0.2
multi-swe-bench	79.4	-0.2
gpqa	79.4	-0.2
vals-legal-research	79.5	-0.2
livebench-data	79.8	+0.2
livebench-language	79.8	+0.2
swe-atlas-qna	79.5	-0.2
vals-code-migration	79.5	-0.1
livebench-coding	79.6	-0.1
livebench-instructions	79.7	+0.1
livebench-math	79.6	-0.1
gdpval	79.6	-0.1
livebench-reasoning	79.6	-0.0
swe-atlas-refactoring	79.6	-0.0
briefcase	79.7	+0.0
livecodebench	79.6	-0.0
apex-agents	79.7	+0.0
aime	79.6	-0.0
terminal-bench-science	79.7	+0.0
vals-excel-modeling	79.6	-0.0
frontiermath	79.6	-0.0
gmmlu	79.6	-0.0
swe-bench-pro	79.7	+0.0
itbench	79.7	-0.0
analyst-agent	79.7	+0.0
math500	79.7	-0.0
enterprise-ops	79.7	-0.0

What evidence is missing from the fit?

Across all collected settings for this model: 39 observations, 16 contributing, 0 matched without graph weight and 23 excluded. Counts do not establish rank eligibility.

1 · Benchmark family or source protocol has not been reviewed
1 · Original SciCode grading defects are under review; retained as supporting evidence, excluded from aggregate and capability scores. Corrected SciCode-Verified requires a separate protocol review. Sources: https://arxiv.org/abs/2608.04975 and https://artificialanalysis.ai/evaluations/scicode
19 · Provider-published claim. Independent source results are admitted separately with exact checkpoint, effort and compatible protocol; this grid does not establish a matched comparison.
1 · Supporting evidence outside the reviewed capability core
1 · No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately

Inspect every observation and exclusion →

Benchmark results

These are published benchmark scores. Supported effort estimates also appear on the Capability leaderboard. Results from different harnesses occupy different rows. Missing cells stay unknown; no scores are copied between effort levels.

Benchmark / source	High
AA-Briefcase EloArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details AA-Briefcase v1.1 is an agentic knowledge work benchmark developed by Artificial Analysis. AA-Briefcase Elo is a combined metric that aggregates rubric pass rate, analytical quality Elo and presentation Elo · Higher is better · Evaluation results measured independently by Artificial Analysis Unit: elo. Source collected 2026-10-01.	1,493.84 Elo Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
AA-LCR v1.1Artificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	79.67% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
AutomationBench-AAArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Share of task objectives completed with no guardrail violations · Higher is better · Benchmark developed by Zapier · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	77.51% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
CritPtArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Benchmark developed by Argonne and UIUC, with contributions from 60+ researchers globally. CritPt is under review by its authors following external feedback · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	27.14% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
GDP.pdf: All-passArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Share of attempts where every atomic criterion passed · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	21.8% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
GDPval-AA v2.1 LeaderboardArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Elo rating for performance on real-world work tasks · Anchored to DeepSeek V4.1 Flash (max) at 1600 · Higher is better · Evaluation results measured independently by Artificial Analysis Unit: elo. Source collected 2026-10-01.	1,611.33 Elo Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
Humanity's Last ExamArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	57.09% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
AA-Omniscience IndexArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct. · Evaluation results measured independently by Artificial Analysis Unit: index. Source collected 2026-10-01.	42.35 Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
AA-Omniscience AccuracyArtificial Analysis · 6000 questions; 1 attempt; correct-answer share; GPT-5.6 Luna medium grader Benchmark details AA-Omniscience Accuracy (higher is better) measures the proportion of correctly answered questions out of all questions, regardless of whether the model chooses to answer · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	49.9% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
SciCodeArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Benchmark curated by scientists across 16 disciplines. SciCode is under review following an independent audit of the dataset · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	61.81% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
Terminal-Bench v4.0Artificial Analysis · mini-SWE-agent v2.4.6; 66 tasks; pass@1 averaged over 3 repeats; 500 steps; task timeout up to 8 hours Benchmark details Benchmark developed by the Laude Institute, Stanford, and open-source contributors · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-01.	57.07% Reported settings Gemini 4 Argon (High); Model source: https://artificialanalysis.ai/models/gemini-4-argon
Vals AI · Finance Agent v2Vals AI · Vals FAB v2 shared six-tool harness; two-hour task limit Benchmark details Overall weighted rubric accuracy, averaged over three runs on 450 tasks. Dealbreaker failures and timeouts score zero. Related categories and all-pass scores are not additional votes. Three-model judge panel; private test set. Unit: percent. Source collected 2026-10-01.	65.4% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":null}
Vals AI · Code Migration v1Vals AI · Vals Code Migration shared agent harness Benchmark details Hidden behavior-test pass rate, equally weighted across 40 repositories (30 CLI and 10 COBOL-to-Java). Anti-cheat failures score zero. Code-quality ratings and language splits are not additional votes. Unit: percent. Source collected 2026-10-01.	68.17% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":null}
Vals AI · Legal Research Bench v1Vals AI · Vals Legal Research shared five-tool harness; three-hour task limit Benchmark details All-pass accuracy on 208 private legal research tasks. Every expert rubric check must pass; GPT-5.4 judges responses. Partial-credit scores and practice-area splits are not additional votes. Unit: percent. Source collected 2026-10-01.	54.81% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":null}
Vals AI · Excel Modeling Benchmark v1Vals AI · Vals EMB shared workbook agent harness; 3.5-hour task limit Benchmark details Mean of seven equally weighted task categories. Template submissions use recalculated Excel cell accuracy; scratch submissions use rubric checks judged by GPT-5.4 Mini XHigh. Partial credit does not mean task completion. Unit: percent. Source collected 2026-10-01.	75.24% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":null}
Vals AI · Vibe Code Bench v1.1Vals AI · Vals modified OpenHands harness; five hours or 1,000 turns per app Benchmark details Mean app test pass rate on the canonical v1.1 test set. Browser Use executes workflow tests; each test requires at least 90% of its substeps. Automated judging can introduce error. Alternate agents are excluded. Unit: percent. Source collected 2026-10-01.	91.91% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":"OpenHands"}
Vals AI · Harvey Legal Agent Benchmark v1Vals AI · Vals Harvey held-out legal file-work protocol; internet disabled Benchmark details Mean task resolution across two judges (GPT-5.5 Medium and Sonnet 4.6), requiring every criterion to pass. Criteria pass rates and practice-area splits do not add votes. Shared Harvey family; Fable 5 fallback run excluded. Unit: percent. Source collected 2026-10-01.	19.58% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":null}
Vals AI · Tax Agent Bench v1Vals AI · Vals Tax Agent Bench shared five-tool harness; three-hour task limit Benchmark details Overall accuracy on 193 private US corporate-tax questions. Each task score is the weighted rubric share, set to zero when a must-pass check fails, then multiplied by (0.7 + 0.3 times citation quality). Timeouts score zero. Category splits and all-pass rates are not additional votes. GPT-5.4 judges expert rubrics. Unit: percent. Source collected 2026-10-01.	76.23% Reported settings {"checkpoint":"gemini-4-argon","evaluatorModel":"google/gemini-4-argon","sourceUpdatedOn":"2026-09-29","reasoning_effort":"high","compute_effort":null,"reasoning":null,"temperature":1,"top_p":null,"max_output_tokens":262144,"verbosity":null,"provider":"Google","harness":null}

What these comparisons mean

Each source reports its own effort labels. Matching labels do not establish equal compute budgets or identical fallback behavior. Open Reported settings to inspect the exact model name, agent and task coverage. Multiple results at one level remain visible rather than selecting the best. Source-reported None or Non-reasoning labels are not treated as evidence that an API supports those settings.

These measurements use their own dated source collection. The leaderboard fits identified configurations jointly; it contains no pooled model ratings. Explicit thinking budgets are separate configurations. Read the methodology.

Download all effort benchmark results · 8,199 results across 246 models · Collected 2026-10-01

Gemini 4 Argon

Compare with

One configuration, seven capabilities

Agentic

Hard reasoning

Coding

Human pref

Knowledge

Multimodal

Long context

Capability score evidence

Benchmark results