Home Model series Claude Haiku Claude Haiku 4.5 τ²-bench Claude Haiku 4.5 · τ²-bench (Retail section) 83.2% · Main text evidence (ev-h45-tau-retail) Claude Haiku 4.5 · τ²-bench (Retail section) 83.2% · Main text evidence (ev-h45-tau-retail) Official source for τ²-bench · Retail section in Claude Haiku 4.5. 发布表报告值 · 83.2%. See “Introducing Claude Haiku 4.5” at τ²-bench Retail 行. Evidence ev-h45-tau-retail.
Source excerpt and location Claude Haiku 4.5 · τ²-bench (Retail section), configuration 发布表报告值, Success rate 83.2% (exact score publication date unconfirmed). See “Introducing Claude Haiku 4.5” at τ²-bench Retail 行. View score conditions
τ²-bench Retail — Haiku 4.5: 83.2%. Official material Introducing Claude Haiku 4.5 · section located Location Main text · τ²-bench Retail 行 Material version 官方发布页;本地保存 HTML/TXT 与主表 PNG:本地留存副本、本地留存副本、本地留存副本;原文发布日期 2025-10-15,当前修订时间未知 Version published 2025-10-15 Disclosure date Unconfirmed · Launch or later disclosure unconfirmed Collected 2026-11-11 Historical basis No reliable historical version evidence available Citation text Claude Haiku 4.5 · τ²-bench (Retail section), configuration 发布表报告值, Success rate 83.2% (exact score publication date unconfirmed). Official source: Introducing Claude Haiku 4.5 (Main text, 官方发布页;本地保存 HTML/TXT 与主表 PNG:本地留存副本、本地留存副本、本地留存副本;原文发布日期 2025-10-15,当前修订时间未知), τ²-bench Retail 行, https://www.anthropic.com/news/claude-haiku-4-5#benchmarks. Benchlog https://benchlog.lekshome.cc/evidence/claude-haiku-4-5/ev-h45-tau-retail/?lang=en
Related scores and conditions Claude Haiku 4.5 · τ²-bench scores and configurations Metric and score Configuration and evaluation conditions Evidence relation Success rate 83.2% Reported 83.2% · Unit % Published Unconfirmed
发布表报告值 Settings SWE-bench: full 500 problems, 50 trials, no test-time compute, 128K thinking budget, default sampling; Terminal-Bench: Terminus 2, 11 runs split between no thinking and 32K thinking; OSWorld: OSWorld-Verified, 100 steps, 4 runs; GPQA/MMMLU/MMMU follow source table. Unknowns AIME 的 Python 工具使用情况按两个分列配置保存;其余各项的逐行配置以主文脚注为准。 Tools Tools allowed Reasoning budget Thinking budget 128K Framework not recorded Protocol τ²-bench Retail。 Conditions 按发布表及模型系统卡中 benchmark 说明;航空策略提示有修订,见报告。 Directly supports this score
View this score in rankings