Atlas

Compare Models

Gemini 3.8 Flash vs Gemini 4 Carbon

Gemini 4 Carbon leads by 7.3 capability points (≈95% interval +4.7 to +9.9): a clear lead over Gemini 3.8 Flash · P(higher) ≈ 100%

Wins 61 of 64 shared benchmarks (2 lost, 1 tied)

Capability vs. relative cost
Gemini 3.8 FlashGemini 4 CarbonClosed weightsOpen weights

Each line runs through one model’s priced reasoning settings, cheapest to most expensive; grey points are other models’ priced settings in the same range. Results whose reasoning effort the source didn’t report are left out wherever a model has named effort levels. Capability and relative cost are the published estimates (relative cost is cost per task as a multiple of GPT-5 (high)); hover a setting, or use the arrow keys, to set it beside the other model’s same setting.

Gemini 3.8 FlashvsGemini 4 Carboncapability vs. relative costData Nov 15, 2026Village LLM Atlasatlas.village.dev

Benchmark wins2 – 61 · 64 shared, 1 tied

Comparing matching setups only (3 benchmarks excluded) · suites counted once
← Gemini 3.8 FlashGemini 4 Carbon →

Gemini 3.8 FlashvsGemini 4 Carbonwins by categoryData Nov 15, 2026Data: Artificial Analysis, Vals AI, Arena and 11 othersVillage LLM Atlasatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Gemini 4 Carbon minus Gemini 3.8 Flash.

ARC-AGI-3 (Semi-Private)General QA35.096.9+61.8
BrokenArXiv 08/2026 (Native Harnesses)Math27.488.1+60.7
BrokenArXiv 09/2026 (Native Harnesses)Math26.782.4+55.7
ArXivMath 08/2026 (Native Harnesses)Math40.493.0+52.6
Terminal-Bench-Science 0.1Science10.562.4+51.9
FrontierSWE V2 Worst@5Code9.556.2+46.7
ArXivMath 09/2026 (Native Harnesses)Math50.096.2+46.2
TxBench-Antibody-Development V1 (60 evaluations)Science15.061.1+46.1
Terminal-Bench 4.0Code19.763.1+43.4
Code Migration v1.1 (Vals Index v2.2 subset)Code38.280.1+41.9
Terminal-Bench 4.0 (Artificial Analysis)Code19.760.6+40.9
Humanity's Last Exam DiamondGeneral QA29.169.1+40.0
Bug Hunt BenchCode17.154.6+37.5
FrontierSWE V2 Best@5Code31.666.7+35.0
ChartographyMultimodal40.972.5+31.6
Artificial Analysis Intelligence Index v5.0.1Indexes30.860.0+29.2
RiemannBenchMath51.280.0+28.8
LMArena Agent Arena Net Improvement (Steering Burden)Agents2.317.0+27.7
Artificial Analysis Intelligence Index v5.0Indexes30.557.7+27.2
Artificial Analysis Coding Agent Index v2.0Indexes33.359.4+26.1
LMArena Code Arena WebDevCode15831768+24.3
AutomationBench 1.0.6 (Zapier)Agents29.753.0+23.3
VIALS-AA (pass^3)Multimodal41.064.0+23.0
AA-Briefcase v1.1 EloAgents12041611+22.3
FrontierCode 2.0 (Artificial Analysis)Code18.740.6+21.8
TxBench Oligo Discovery V1 (113 evaluations)Science35.155.8+20.7
Blueprint-Bench 2Multimodal0.40.6+20.0
MKA (Artificial Analysis)Professional Work35.354.3+19.1
LMArena Agent Arena Praise vs ComplaintAgents4.941.0+18.1
APEX-Agents 1.1Professional Work72.990.4+17.5
CritPt v2 (agentic, avg@5)Science65.182.3+17.1
AutomationBench-AA 1.0.6Agents60.977.8+16.9
Legal Research BenchLegal38.955.8+16.8
TxBench Antibody Discovery V1 (100 evaluations)Science49.065.0+16.0
VIALS-AA (average of 3)Multimodal52.067.3+15.3
AA-Omniscience v1.1 IndexIndexes23.753.5+14.9
GDPval-AA v2.1Professional Work14421733+12.8
BuseyBench SVGMultimodal6.67.9+12.3
APEX-Accounting 1.1Finance & Accounting58.870.7+11.9
Harvey LAB (Mercor run, undisclosed ~213-task cohort)Legal11.423.0+11.6
TxBench Preclinical Pharmacology V1 Clean RerunsScience54.766.0+11.3
Artificial Analysis Progress Index v1Indexes320351+10.7
Vibe Code Bench v1.1Code78.789.1+10.4
LMArena Agent Arena Steering BurdenAgents-1.319.1+10.2
Tax Agent BenchFinance & Accounting66.876.4+9.6
LMArena Agent Arena Confirmed SuccessAgents1.318.4+8.5
Vals RSI Index v1.1 (four tasks)Agents19.927.7+7.8
LMArena Agent Arena Bash ReliabilityAgents9.9-5.2−7.5
τ³-Banking (Mercor run, bm25_grep retrieval)Agents48.855.7+6.9
VIALS-AA (pass@3)Multimodal64.070.8+6.8
GDP.pdf (Artificial Analysis, All-pass)Professional Work22.829.6+6.8
Vals Finance IndexIndexes60.567.0+6.6
DAXBench v2Code87.293.6+6.4
Vals Legal IndexIndexes69.875.9+6.1
HLAB-Vals-1.1Legal12.117.5+5.4
ARC-AGI-2General QA89.294.2+5.0
LMArena Text Arena EloChat & Writing14971526+5.0
MMMU ProMultimodal84.589.2+4.7
AA-LCR v1.1Search84.079.3−4.7
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)Health & Medicine15.619.4+3.9
Excel Modeling Benchmark v1.1Finance & Accounting72.976.5+3.6
Finance Agent v2.1Finance & Accounting65.867.2+1.4
LMArena Agent Arena Tool HallucinationAgents0.40.50.0
ARC-AGI-1 Semi-PrivateGeneral QA98.598.50.0

Gemini 3.8 FlashvsGemini 4 Carbonnormalised score marginsData Nov 15, 2026Data: Artificial Analysis, Vals AI, Arena and 11 othersVillage LLM Atlasatlas.village.dev

Win counts use each benchmark’s declared score direction. Represented suite slices and the derived Capability score are excluded from headline totals. Missing or incompatible setups are excluded, never counted as losses.
Loading Atlas data…