BenchLeader
Menu
LeaderboardsModelsProvidersBenchmarksCompareReleasesBlogMethodology
Anthropic

Claude 3.5 Sonnet

Best configuration ranks #396 of 610 on the BenchLeader Index at 45.1 ±3.7. Last measured 2 Sept 2026. Released 22 Oct 2024.

Blended price
$6.00/M
$3.00 in · $15.00 out
Output speed
–
First answer
–
Context
200k
How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.
  1. Overall index45
  2. Reasoning42
  3. Coding44
  4. Agents & tools43
  5. Maths40
  6. Knowledge47
  7. Human preference55
  8. Multimodal46
  9. Composite39

Benchmark results

One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Reasoning

BenchmarkdefaultSource
GPQA Diamond55.3%#190Epoch AI Benchmarking Hub
Humanity's Last Exam4.1%#42Scale AI / CAIS
SimpleBench41.4%#57SimpleBench
LMArena Hard Prompts1398#154LMArena
GPQA Diamond (AA)not in index59.9%#358Artificial Analysis
Humanity's Last Exam (AA)not in index3.7%#490Artificial Analysis
GPQA Diamond (Vals)not in index59.3%#106Vals AI

Coding

BenchmarkdefaultSource
WeirdML40.0%#101WeirdML
GSO-Bench4.6%#24GSO-Bench
LMArena Coding1435#143LMArena
LiveCodeBench49.6%#112Vals AI
Aider Polyglot51.6%#24Aider polyglot leaderboard
SWE-bench Verified (any scaffold)not in index63.4%#26SWE-bench

Agents & tools

BenchmarkdefaultSource
Cybench17.5%#12Cybench

Maths

BenchmarkdefaultSource
OTIS Mock AIME8.5%#211Epoch AI Benchmarking Hub
MATH Level 557.0%#47Epoch AI Benchmarking Hub
AIME (Vals)10.0%#86Vals AI
MGSM92.6%#15Vals AI

Knowledge

BenchmarkdefaultSource
MMLU-Pro78.4%#99Vals AI
CorpFin53.6%#94Vals AI
TaxEval70.2%#91Vals AI
MedQA83.2%#62Vals AI

Human preference

BenchmarkdefaultSource
LMArena Text1374#158LMArena

Multimodal

BenchmarkdefaultSource
LMArena Vision1125#103LMArena
VISTA38.7%#40Scale AI SEAL
MMMU (validation)68.3%#23MMMU
MMMU-Pro (official)not in index51.5%#16MMMU

Composite

BenchmarkdefaultSource
Epoch Capabilities Indexnot in index133.6#123Epoch AI Benchmarking Hub
AA Intelligence Index7.9#375Artificial Analysis

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache. Reasoning tokens are not modelled.

WorkloadTokens in / outCostWith cachingTime
Chat reply400 / 300$0.0057––
Summarise a 30-page report12,000 / 600$0.045––
Code edit6,000 / 1,500$0.041––
Agentic coding session60,000 / 4,000$0.240––
Structured extraction2,000 / 200$0.0090––

See also

Best for codingAll Anthropic models

Data as of 9 Sept 2026. Compare with another model.