Unless otherwise noted: adaptive thinking at max effort, default temperature/top_p, five-trial average; context window is evaluation-dependent and no more than 1M tokens.
未知项
每个 benchmark 的具体 context window 未逐项列明。
工具
工具条件未记录
推理预算
max
框架
mini-swe-agent
协议
ProgramBench; hidden-test pass rate on 166 retained tasks using mini-swe-agent harness.
条件
34 flaky tasks were excluded from the original 200; internet and decompilation tools were not used; upstream six-hour limit was omitted; task contexts reached up to 1M tokens.
Claude Haiku 5.5 · ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分 · Scores and configurations
Claude Haiku 5.5 record for ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分. 1 score record for this release, including 1 exact value (may include historical records). Includes 1 source location.
Claude Haiku 5.5 · ProgramBench scores and configurations
Metric and score
Configuration and evaluation conditions
Sources
Hidden test pass rate82%Reported 82.0% · Unit %
Published Unconfirmed
System Card 标准 max effort
Settings
Unless otherwise noted: adaptive thinking at max effort, default temperature/top_p, five-trial average; context window is evaluation-dependent and no more than 1M tokens.
Unknowns
每个 benchmark 的具体 context window 未逐项列明。
Tools
Tool access not recorded
Reasoning budget
max
Framework
mini-swe-agent
Protocol
ProgramBench; hidden-test pass rate on 166 retained tasks using mini-swe-agent harness.
Conditions
34 flaky tasks were excluded from the original 200; internet and decompilation tools were not used; upstream six-hour limit was omitted; task contexts reached up to 1M tokens.