Terminal-Bench

DeepSeek V4-Flash 0731通过微调实现性能跃升，成本降低60%，但未改变架构。

DeepSeek V4 Flash在AI Gateway更新权重后Terminal-Bench得分提升至82.7，增强代理能力且无需代码修改。

OpenAI 推出 GPT 5.6 系列模型，包含 Soul、Terror 和 Luna，Soul 超越 Claude Mythos 5 在终端任务表现。

Fireworks AI 已上线 GLM 5.2 模型，支持 1M-token 上下文，专注于代码生成，并在多个基准测试中表现优异。

Google releases the Gemini 3.5 family, starting with 3.5 Flash for complex agentic workflows. It outperforms 3.1 Pro on coding and agent benchmarks and runs 4x faster, reaching 12x in Antigravity.

Thomas Wolf is excited about the extension of Terminal-Bench to scientific fields, known as Terminal-Bench Science. This benchmark evaluates AI models' ability to control tools via the command line to achieve scientific goals. It's open for contributions of real scientific workflows until August 2026, aiming to improve AI models' assistance in research work.

A harness is the core infrastructure for building AI Agents, consisting of tools, execution environments, system prompts, and file systems. By optimizing harness engineering, developers can significantly boost Agent performance on benchmarks like Terminal Bench without changing the underlying model.