Claude Haiku 5.5 · Terminal-Bench(4.0) 39.2% · 主文证据(ev-h55-terminal)
Claude Haiku 5.5 的 Terminal-Bench · 4.0 官方出处。发布表 Max effort · 39.2%。见《Introducing Claude Haiku 5.5》本地留存副本 lines 32-33; official HTML table。证据 ev-h55-terminal。
Claude Haiku 5.5 · Terminal-Bench(4.0),发布表 Max effort配置,Score 39.2%(具体成绩公布时间未确认)。见《Introducing Claude Haiku 5.5》本地留存副本 lines 32-33; official HTML table。 查看成绩条件
Claude Haiku 5.5 · Terminal-Bench(4.0),发布表 Max effort配置,Score 39.2%(具体成绩公布时间未确认)。官方出处 Introducing Claude Haiku 5.5(主文,官方发布页;本次保存 HTML/TXT:本地留存副本、current-haiku-5-5-main.txt;发布日期 2026-10-07,当前修订时间未知)本地留存副本 lines 32-33; official HTML table,https://www.anthropic.com/news/claude-haiku-5-5#performance。Benchlog https://benchlog.lekshome.cc/evidence/claude-haiku-5-5/ev-h55-terminal/
关联成绩与条件
Claude Haiku 5.5 · Terminal-Bench 成绩与配置
指标与成绩
配置与评测条件
证据关系
Score39.2%原文 39.2% · 单位 %
公布 未确认
发布表 Max effort
设置
Anthropic 主文表格;unless stated otherwise, adaptive thinking and max effort。
Terminal-Bench 4.0; 66 tasks; ten trials per task (660 trials); Claude Code --bare at max thinking effort.
条件
Safeguards enabled without fallback; requests were flagged in 1.8% of trials (12/660), causing failure. No internet egress; historical resources were pre-cached. The card estimates this did not move the result outside noise but did not rigorously measure the effect.
Claude Haiku 5.5 · Terminal-Bench (4.0) 39.2% · Main text evidence (ev-h55-terminal)
Official source for Terminal-Bench · 4.0 in Claude Haiku 5.5. 发布表 Max effort · 39.2%. See “Introducing Claude Haiku 5.5” at 本地留存副本 lines 32-33; official HTML table. Evidence ev-h55-terminal.
Claude Haiku 5.5 · Terminal-Bench (4.0), configuration 发布表 Max effort, Score 39.2% (exact score publication date unconfirmed). See “Introducing Claude Haiku 5.5” at 本地留存副本 lines 32-33; official HTML table. View score conditions
Terminal-Bench 4.0; 66 tasks; ten trials per task (660 trials); Claude Code --bare at max thinking effort.
Conditions
Safeguards enabled without fallback; requests were flagged in 1.8% of trials (12/660), causing failure. No internet egress; historical resources were pre-cached. The card estimates this did not move the result outside noise but did not rigorously measure the effect.