From c633508c94c1d0ccff1dcf53878193f759e006d5 Mon Sep 17 00:00:00 2001 From: Eric Lee Date: Wed, 29 Jul 2026 00:31:24 -0700 Subject: [PATCH] docs(readme): add Terminal-Bench 2.1 headline banner (#766) Lead the README with a centered Terminal-Bench 2.1 banner (80.9% on claude-opus-5, would slot around 3rd on the public board) alongside the existing /eco and DeepSeek headline banners; add a matching headline section to the ZH README. Co-Authored-By: Claude Opus 4.8 --- README.md | 16 ++++++++++++++++ docs/i18n/README_ZH.md | 16 ++++++++++++++++ 2 files changed, 32 insertions(+) diff --git a/README.md b/README.md index f46ed117..0f1ce9dd 100644 --- a/README.md +++ b/README.md @@ -26,6 +26,22 @@
+# 🏆 Terminal-Bench 2.1 + +# **80.9%** on Claude Opus 5 — a top-tier open-source result + +### 72 of 89 tasks solved (pass@1) on the verified 89-task suite. On the [public leaderboard](https://www.tbench.ai/leaderboard/terminal-bench/2.1) that slots **around 3rd** — behind Claude Code / Fable 5 (83.8%) and Codex / GPT-5.5 (83.1%), **ahead of Claude Code on Opus 4.8 (78.9%)**. + +ClawCodex — the open-source Python rebuild of Claude Code — run headless on `claude-opus-5` at `effort=xhigh`. +A single run (k=1) against the board's k=5 averages; benchmarked on `main` at #756. Fully reproducible from the +open-source [Harbor adapter](eval/harbor/). **[Read the full breakdown ↓](#-news)** + +
+ +*** + +
+ # 🌿✂️ /eco Token Compression # Same session, **80% fewer** Bash-output tokens diff --git a/docs/i18n/README_ZH.md b/docs/i18n/README_ZH.md index 347dbe58..5b649ac4 100644 --- a/docs/i18n/README_ZH.md +++ b/docs/i18n/README_ZH.md @@ -94,6 +94,22 @@ clawcodex --dangerously-skip-permissions # 启动 REPL *** +## 🏆 Terminal-Bench 2.1 —— Opus 5 上 **80.9%**，顶尖开源成绩 + +在经过验证的 89 任务 **Terminal-Bench 2.1** 套件上，ClawCodex 以 `claude-opus-5`（`effort=xhigh`）无头运行，解决了 **72 / 89** 个任务（**80.9% pass@1**，单次运行）。在[公开排行榜](https://www.tbench.ai/leaderboard/terminal-bench/2.1)（k=5 平均）上这大约排在**第 3 左右**： + +| Agent | 模型 | 准确率 | +|---|---|---:| +| Claude Code | Fable 5 | 83.8% | +| Codex | GPT-5.5 | 83.1% | +| **ClawCodex** | **Opus 5** | **80.9%** | +| Terminus 2 | Fable 5 | 80.4% | +| Claude Code | Opus 4.8 | 78.9% | + +落后于 Claude Code / Fable 5 与 Codex / GPT-5.5，**领先 Claude Code 搭配 Opus 4.8（78.9%）**。单次 k=1 对照榜单的 k=5 平均值，基准运行在 v1.3.0 打标签前的 `main`（#756），完整可由开源 [Harbor 适配器](../../eval/harbor/) 复现。 + +*** + ## 🏆 SWE-bench Verified —— 相同模型下 `clawcodex` 超越 `openclaude` ![SWE-bench Verified —— clawcodex vs openclaude on Gemini 2.5 Pro](../../assets/swebench-verified-gemini.png)