benchgap
Model compare

Claude Opus 5.5 vs Gemini 4 Carbon

Claude Opus 5.5 stands at 94.8 on 182 benchmarks, 89 measured (rank #8 of 394), Gemini 4 Carbon at 96.3 on 128 benchmarks, 37 measured (rank #5 of 394) - a frontier model. Of the 128 benchmarks both have a score on, Claude Opus 5.5 leads on 12 and Gemini 4 Carbon on 21 where both are measured.

BenchmarkClaude Opus 5.5Gemini 4 CarbonΔ (pp)
AA-GPQA Diamond96.2%94.5%+1.7
AA-HLE61.4%52.8%+8.6
AA-SciCode66.9%67.7%-0.8
AIME2698.1%98.1%+0.0
Artificial Analysis Intelligence Index59.1%60.0%-0.9
CritPt31.7%30.7%+1.0
GDPval-AA68.3%62.7%+5.6
GPQA Diamond96.3%95.5%+0.9
HLE63.8%72.6%-8.8
AA-LCR84.7%79.3%+5.4
AA-Omniscience Accuracy65.5%58.3%+7.2
Terminal-Bench 2.189.1%89.1%-0.0
BrowseComp85.3%96.6%-11.3
HLE w/ tools67.7%66.9%+0.8
LongBench v270.0%65.1%+4.9
AA Coding Index85.1%86.6%-1.5
APEX-Agents37.5%69.4%-31.9
BioMysteryBench (human-difficult)50.0%49.0%+1.0
AutomationBench40.0%53.9%-13.9
CyberGym84.4%87.9%-3.5
JobBench63.1%73.0%-9.9
SWE-bench Pro89.9%92.1%-2.2
AA Agentic Index69.5%60.0%+9.5
MMLU-Pro86.1%85.5%+0.5
FailureSensorIQ67.9%68.7%-0.8
GPQA96.3%96.5%-0.1
Vals GPQA Diamond98.8%93.8%+5.1
Vals LiveCodeBench89.3%94.8%-5.6
Vals MMLU-Pro92.0%89.1%+2.9
Chess Puzzles67.0%78.0%-11.0
AA MMLU-Pro91.1%90.0%+1.1
AA-AnalystAgent56.3%55.3%+1.0
AA Tau3 Banking100.0%50.6%+49.5
ApprenticeBench42.6%49.5%-6.8
ARC-AGI-197.5%97.5%+0.0
ARC-AGI-291.7%92.1%-0.4
BabaIsBench100.0%100.0%+0.0
CursorBench 3.270.5%77.4%-7.0
EBR-bench71.4%87.6%-16.2
Terminal-Bench 3.050.4%49.6%+0.8
FrontierCode 1.1 Main54.4%53.6%+0.8
FrontierMath v2 (Tier 4)95.0%95.1%-0.1
FrontierMath v2 (Tiers 1-3)91.2%90.5%+0.7
FrontierSWE v262.3%61.2%+1.1
Furniture Assembly83.3%77.5%+5.8
MaxxBench76.4%76.2%+0.2
MazeBench13.0%16.0%-3.0
MirrorCode95.6%70.4%+25.2
Mystery Game Puzzles71.0%73.0%-2.0
NYT Connections88.5%99.3%-10.8
Terminal-Bench 4.165.8%64.6%+1.2
Terminal-Bench Hard70.7%70.4%+0.3
Vals SWE-bench96.9%100.0%-3.1
Terminal-Bench 2.1 (Vals AI)87.6%84.6%+3.0
VulcanBench v387.1%87.6%-0.6
CWE-Bench-AA58.3%63.3%-5.0
AA Cyber Index28.7%46.3%-17.6
DeepsecBench-AA26.9%45.8%-18.9
Harvey LAB-AA v1.14.2%9.6%-5.4
HealthBench Professional (length-adjusted, Artificial Analysis)59.9%64.1%-4.2
HealthBench Professional (raw, Artificial Analysis)71.4%71.0%+0.4
Bug Hunt Bench41.7%57.3%-15.6
Chartography (no tools)64.4%72.5%-8.1
Chartography (tools)89.0%91.3%-2.3
CursorBench 4.057.8%60.4%-2.6
CWE-bench v167.0%71.0%-4.0
DeepSWE74.2%75.6%-1.4
DRACO87.4%87.9%-0.5
FrontierCode 2.047.3%42.6%+4.7
FrontierPhysics27.4%19.0%+8.4
HLE w/o tools64.4%63.9%+0.5
OSWorld 2.181.8%78.6%+3.2
OSWorld 2.1 Offline87.2%81.5%+5.7
PostTrainBench v1.149.3%49.3%-0.0
SpireClimb5.8%38.2%-32.4
StealthMark72.6%74.8%-2.2
SWE-bench Multimodal61.4%74.1%-12.7
Terminal-Bench 4.066.4%67.5%-1.1
Terminal-Bench-Science 0.158.7%66.3%-7.6
Toolathlon-Verified77.8%77.8%-0.0
VulcanBench Frontier v491.1%92.7%-1.6
AA AutomationBench69.5%77.8%-8.3
GDP.pdf26.2%28.4%-2.2
AA Terminal-Bench 4.059.6%65.2%-5.6
ARC-AGI-338.8%63.9%-25.0
ARC-AGI-3 Provider Adapter97.3%96.9%+0.4
cursorBench4158.1%61.5%-3.4
CWE-bench67.1%67.3%-0.2
MKA53.0%54.3%-1.3
SkillsBench official leaderboard (BenchFlow)58.7%60.3%-1.6
BinaryAudit35.8%35.8%+0.0
OTelBench22.5%22.5%+0.0
FACTS Multimodal56.4%60.2%-3.8
Hermes Bench76.7%75.4%+1.4
Hermes Index63.3%91.7%-28.4
ExploitBench100.0%100.0%+0.0
ExploitGym33.3%53.5%-20.2
HMMT Feb 2026100.0%100.0%+0.0
HMMT Nov 202596.1%96.1%+0.0
MMAnswerBench97.1%96.9%+0.2
MMLU-Redux95.9%98.2%-2.3
NL2Repo62.6%73.8%-11.2
PostTrainBench v147.7%44.8%+2.9
SuperGPQA100.0%100.0%+0.0
PostTrainBench43.8%41.0%+2.8
DAYJOB24.5%19.7%+4.8
DAYJOB Healthcare25.2%15.6%+9.6
Finance Agent v262.3%62.3%+0.0
FrontierCode 1.1 Extended63.6%62.5%+1.1
HealthBench (raw)68.1%64.1%+4.0
HealthBench (length-adjusted)60.6%58.4%+2.2
HealthBench Professional65.6%64.6%+1.0
HealthBench Professional (raw)77.1%66.5%+10.6
HLE-Verified54.9%55.0%-0.1
LABBench284.6%84.6%+0.0
ScarfBench70.5%69.8%+0.7
BlenderWorld (text-only)86.1%85.6%+0.5
Cookie-Bench91.5%90.5%+1.0
STEMBench2.6%2.7%-0.0
Agents' Last Exam51.7%51.7%+0.0
OpenHarmony Bench v1.263.8%63.8%+0.0
SEC-Bench Pro85.0%88.2%-3.2
MRCRv286.0%85.7%+0.3
MMLongBench-Doc61.6%61.4%+0.2
SWE-Marathon57.3%67.8%-10.5
BioBasicsBench69.0%68.1%+0.9
LifeVal83.4%83.5%-0.0
ExploitBench Internal Port18.8%53.4%-34.6

General performance is the share of benchmark duels a model is expected to win against every other ranked model: a Bradley-Terry fit over every head-to-head result on the listed benchmarks, measured and estimated, each pair held against each other on the benchmarks both have a score on. A model measured only on benchmarks apart from the main field has no standing. The frontier is the strongest tenth. Every estimate carries its error and confidence on the model pages.