Anthropic · proprietary

Claude Sonnet 4.5

Claude Sonnet 4.5 by Anthropic appears in 2 sources with Reasoning at 40.2. Best read for Code quality, Coding, LLM.

CreatorAnthropic

Release date2025-09-29

Knowledge cutoffNot published

Context200K tokens

Input price$3/M tokens

Output price$15/M tokens

Modalitytext + vision

CountryUS

Metrics

All source-backed metrics

Open in leaderboard

LLM Stats Rank16ranking · source_rank

Reasoning40.2reasoning · index_reasoning

Math33.26math · index_math

Coding31.46coding · index_code

Writing35.23writing · index_communication

Vision18.73multimodal · index_vision

Tool calling33.59tool_calling · index_tool_calling

Long context19.19long_context · index_long_context

Healthcare16.6domain · index_healthcare

GPQA83.4 %reasoning · gpqa_score

AIME 202587 %math · aime_2025_score

Code Arena1,412.14 %coding · coding_arena_score

MMMLU89.1 %reasoning · mmmlu_score

OSWorld61.4 %agent · osworld_score

Terminal Bench50 %coding · terminal_bench_score

TAU2 Retail86.2 %agent · tau_bench_retail_score

Context200,000 tokenscontext · context

Speed117.59 c/sperformance · throughput

Latency932.75 msperformance · latency

Input price3 $/Mpricing · input_price

Output price15 $/Mpricing · output_price

SWE-Bench Verified82 %metric

Evidence

Citations and source overlap

VellumTop models per benchmark from page-level structured data (JSON-LD).Retrieved 2026-05-20 LLM StatsPublic full LLM leaderboard advanced explorer snapshot, including ranking, benchmark indexes, pricing, speed, context, license, and model metadata.Retrieved 2026-06-01

FAQ

How should I read this profile?

Treat this as a source-backed model dossier, not an EvalKit-run verification. The public values are replicated from linked sources and kept source-scoped.

Is Claude Sonnet 4.5 verified by EvalKit?

No. EvalKit currently shows 0 verified rows until real run evidence exists.

Why can metrics disagree?

Different sources test different tasks, dates, prompts, and aggregation methods. EvalKit keeps those differences visible instead of merging them into a fake universal score.