Claude Opus 5 vs Claude Sonnet 5.5 vs Claude Opus 5.5

Best cell per row highlighted. Null means undisclosed — never counted as zero.

Anthropic
Claude Opus 5GA
9.5
AI Panel
Anthropic
Claude Sonnet 5.5GA
9.3
AI Panel
Anthropic
Claude Opus 5.5GA
9.7
AI Panel

Identity & lifecycle

ProviderAnthropicAnthropicAnthropic
Family / tierClaude 5 · OpusClaude 5 · SonnetClaude 5 · Opus
StatusGAGAGA
Released2026-07-242026-09-282026-09-22
Knowledge cutoff2026-052026-062026-06

Architecture & context

Context window1M1M1M
Max output tokens128K128K128K
Input modalitiestext, imagetext, imagetext, image
Reasoning modeoptionaloptionalalways
Open weightsNoNoNo

Pricing (per Mtok)

Input$5$2$4
Output$25$10$20
Cached input$0.5$0.2$0.2
Batch input$2.5$1$2
Free tierNoYesNo

Speed

Output speed (tok/s)—97.672.8
Time to first token (s)—12.723.7
Latency tiermediumfastmedium

Trust & deployment

Trains on inputsNoNoNo
LicenseProprietaryProprietaryProprietary
CloudsBedrock, Vertex AI, Azure AI FoundryBedrock, Vertex AI, Azure AI FoundryBedrock, Vertex AI, Azure AI Foundry
ComplianceSOC2, ISO27001, ISO42001, HIPAA, CSA-STARSOC2, HIPAA, ISO27001, ISO42001SOC2, ISO27001, ISO42001, HIPAA

AI Panel scoring

Unified score9.59.39.7
Decision Maker9.59.59.5
Domain Strategist99.29.5
Finance Lead8.58.89
Domain Practitioner9.59.39
Power User99.29.5
Skeptic888.5
Value score8.59.29

Benchmarks

SWE-bench Verified96——
Humanity's Last Exam56.356.964.4
LMArena Elo149314711504
LMArena Coding Elo153315361538
Artificial Analysis Index—5658

Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.