BenchLeader
Menu
Best LLMsChoose a modelCoding plansLeaderboardsModelsBenchmarksProvidersReleases9

Compare

Pick up to six models. You get a plain-language verdict, a spec sheet with the best value in each row highlighted, a chart, and every benchmark score side by side.

Claude Opus 5.5maxGemini 4 Carbonhigh
Drag a model to change the order.
Verdict
  • Claude Opus 5.5 (max) and Gemini 4 Carbon (high) are level on quality (69.1 vs 68.8).
  • Claude Opus 5.5 (max) is strongest in coding, long context, maths, multimodal.
  • Gemini 4 Carbon (high) is strongest in agents & tools, knowledge, reasoning, human preference.
  • They cost about the same ($8.00 per 1M blended).
  • Gemini 4 Carbon (high) starts answering in 1.6 min; Claude Opus 5.5 (max) takes 10.5 min.
MetricClaude Opus 5.5
max Anthropic
Gemini 4 Carbon
high Google
Quality
BenchLeader Index50 = average model69.1 — best of the models compared68.8
Reasoning72.977.8 — best of the models compared
Coding80.2 — best of the models compared74.4
Agents & tools63.078.3 — best of the models compared
Maths74.0 — best of the models compared70.9
Knowledge74.176.1 — best of the models compared
Human preference–72.8
Multimodal70.6–
Long context67.3 — best of the models compared64.5
Speed and latency
Output speed97 tok/s — best of the models compared85 tok/s
Time to first token3.53 s–
Time to first answerafter any reasoning632.8 s97.9 s — best of the models compared
Context window1M — best of the models compared1M — best of the models compared
Price
Input, per 1M tokens$4.00 — best of the models compared$4.00 — best of the models compared
Output, per 1M tokens$20.00 — best of the models compared$20.00 — best of the models compared
Blended (3 in : 1 out)$8.00 — best of the models compared$8.00 — best of the models compared
Cached input, per 1Mprompt-cache read price$0.200–
Index per dollar8.6 — best of the models compared8.6 — best of the models compared
Cost and time per task
Chat reply: cost400 in / 300 out$0.0076 — best of the models compared$0.0076 — best of the models compared
Chat reply: cost with caching75% cache hits$0.0065–
Chat reply: time10.6 min1.7 min — best of the models compared
Summarize a 30-page report: cost12,000 in / 600 out$0.060 — best of the models compared$0.060 — best of the models compared
Summarize a 30-page report: cost with caching75% cache hits$0.026–
Summarize a 30-page report: time10.7 min1.8 min — best of the models compared
Code edit: cost6,000 in / 1,500 out$0.054 — best of the models compared$0.054 — best of the models compared
Code edit: cost with caching75% cache hits$0.037–
Code edit: time10.8 min1.9 min — best of the models compared
Agentic coding session: cost60,000 in / 4,000 out$0.320 — best of the models compared$0.320 — best of the models compared
Agentic coding session: cost with caching75% cache hits$0.149–
Agentic coding session: time11.2 min2.4 min — best of the models compared
Structured extraction: cost2,000 in / 200 out$0.012 — best of the models compared$0.012 — best of the models compared
Structured extraction: cost with caching75% cache hits$0.0063–
Structured extraction: time10.6 min1.7 min — best of the models compared

Chart

Quality againstYour models are ringed and labelled; the rest is the frontier.
40506070$0.100$0.200$0.500$1.00$2.00$5.00$10.00$20.00Blended price, $ per 1M tokens (log)BenchLeader IndexClaude Opus 5.5 (max)Claude Opus 5.5 (max)Gemini 4 Carbon (high)Gemini 4 Carbon (high)GPT-6 AstraGPT-6.1 Sol (max)Claude Fable 5.1 (max)Claude Fable 5Gemini 4 Argon (high)GPT-5.6 Sol (max)Claude Opus 5 (max)GPT-6.1 Terra (max)Gemini 3.7 Flash (high)Kimi K3.1 ProDeepSeek V4.1 Flash 1017 (max)GPT-5.6 Luna (xhigh)DeepSeek V4 Flash (high)GLM 5.3 FlashClaude Haiku 5.5 (xhigh)Step 3.5 FlashLing 3.0 FlashMistral Small 3.2
  • Anthropic5
  • OpenAI5
  • Google3
  • DeepSeek2
  • Ant Group1
  • Mistral AI1
  • Moonshot AI1
  • StepFun1
  • Zhipu AI1
Dashed line: the frontier. Purple-ringed points are not beaten on both axes by any other model shown.

Every benchmark

Raw scores side by side, grouped by category. Best value in each row is purple. “Ref.” benchmarks are shown but not counted in the index.

Benchmark Claude Opus 5.5 max Gemini 4 Carbon high
Reasoning
GPQA Diamond90.6%–
Arena Hard Prompts–1548
LiveBench 1.0 Reasoning (ref.)92.2%–
Humanity's Last Exam (AA) (ref.)61.4% — best of the models compared60.4%
HLE-Diamond (AA)67.9%69.1% — best of the models compared
ARC-AGI-197.5%–
ARC-AGI-291.7%–
CritPt (ref.)31.7%36.0% — best of the models compared
CritPt v2 hard subset (AA)61.0%63.0% — best of the models compared
Chess Puzzles (ref.)67.0%–
Mystery Game Puzzles (ref.)71.0%–
LMCA (ref.)68.2%–
DTBench (ref.)98.9%–
Coding
SciCode66.9% — best of the models compared64.2%
Arena Coding–1573
Arena WebDev1810 — best of the models compared1768
LiveBench 1.0 Coding (ref.)89.3%–
SciCode (AA) (ref.)66.9% — best of the models compared64.2%
Vibe Code Bench v1.1 (ref.)–89.1%
CursorBench (ref.)57.8%–
FrontierSWE (ref.)62.3%–
Terminal-Bench 4.0 (AA) (ref.)59.6%65.2% — best of the models compared
FrontierCode 1.1 (AA) (ref.)52.2% — best of the models compared48.5%
FrontierCode 2.0 (AA)46.4% — best of the models compared40.6%
Agents & tools
Terminal-Bench64.8%–
APEX-Agents73.5%–
Arena Agent20.9 — best of the models compared16
LiveBench Agentic Coding (ref.)71.7%–
GDPval-AA v2.1, normalised score (ref.)68.3%–
GDPval-AA v2.1 (ref.)1866 — best of the models compared1733
ITBench SRE (AA) (ref.)38.2%–
Analyst Agent (AA) (ref.)56.3%–
CUA-bench (ref.)14.0%–
Finance Agent v2.1 (ref.)–67.2%
Harvey's Legal Agent Benchmark v1.1 (ref.)–17.5%
Legal Research Bench (ref.)–55.8%
Robo-bench (ref.)14.6%–
Tax Agent Bench (ref.)–48.9%
Terminal-Bench 4.0 (Vals) (ref.)–63.1%
AutomationBench (ref.)69.5%77.8% — best of the models compared
GDP.pdf (AA) (ref.)26.2%28.4% — best of the models compared
Harvey LAB-AA v1.1 (ref.)4.2%–
AA-Briefcase v1.1 (ref.)1800 — best of the models compared1609
Terminal-Bench-Science 0.1 (AA) (ref.)59.0%64.8% — best of the models compared
CWE-Bench (AA)58.3%–
DeepSec-Bench (AA)26.9%–
CyberGym end-to-end (AA)0.8%–
Maths
FrontierMath Tiers 1–391.2%–
FrontierMath Tier 495.0%–
OTIS Mock AIME100.0%–
ProofBench100.0%–
LiveBench Mathematics (ref.)97.1%–
Arena Maths–1525
Knowledge
SimpleQA Verified72.2%–
LiveBench 1.0 Data Analysis (ref.)80.3%–
AA-Omniscience v1.148.153.5 — best of the models compared
Excel Modeling Benchmark v1.1 (ref.)–76.5%
EBR-bench (ref.)71.4%–
GDP.pdf (ref.)30.6%–
AA-Omniscience v1.1: accuracy (ref.)65.5% — best of the models compared58.3%
AA-Omniscience v1.1: non-hallucination (ref.)49.7%88.6% — best of the models compared
HealthBench Professional (AA) (ref.)59.9%–
Instruction following
LiveBench Language (ref.)86.3%–
LiveBench Instruction Following (ref.)65.7%–
Arena Instruction Following (ref.)–1525
Human preference
Arena Text–1528
Arena Creative Writing (ref.)–1515
Arena Multi-turn (ref.)–1559
Arena Longer Queries (ref.)–1542
Multimodal
MMMU-Pro87.7%–
Code2Video-AA (ref.)1536–
AA-3D (ref.)1196–
Long context
AA-LCR v1.184.7% — best of the models compared79.3%
MLCR (ref.)27.8%–
Composite
LiveBench 1.0 (ref.)83.2%–
AA Intelligence Index v5.0 (ref.)58.459.4 — best of the models compared
AA Intelligence Index v4.3.2 (ref.)57.6–
AA Progress Index v1 (ref.)350351.6 — best of the models compared
Vals Index (ref.)–70.4
Vals Finance Index (ref.)–66.3%
Vals RSI Index (ref.)37.3%–
AA Cyber Index (ref.)28.7–
Safety & honesty
FORTRESS (ref.)5.8%–

Demand pairs first, then every pairing of the top 15 models: each with a verdict, category wins, price and speed. Each model’s page links the rest of its comparisons.