Atlas

Benchmarks

← All benchmarks

Long-Horizon Terminal-Bench 1.0 (solved, reward ≥ 0.95)

Code · 2026-07-09

The share of the 46 Long-Horizon Terminal-Bench 1.0 tasks an agent solves, counting a task when its hidden verifier awards a reward of at least 0.95 out of 1, in percent. It comes from the same single 90-minute Terminus 2 attempt per task as the mean-reward headline, which keeps the partial credit this threshold discards.

Benchmark source: arXiv

Top models (higher is better)

ModelScoreSource
Grok 4.528.3LHTB 1.0 leaderboard (Grok 4.5) · Tencent
Claude Fable 526.1LHTB 1.0 leaderboard (Claude Fable 5) · Tencent
Opus 4.819.6LHTB 1.0 leaderboard (Claude Opus 4.8) · Tencent
Sonnet 517.4LHTB 1.0 leaderboard (Claude Sonnet 5) · Tencent
Gemini 3.6 Flash15.2LHTB 1.0 leaderboard (Gemini 3.6 Flash) · Tencent
GPT-5.515.2LHTB 1.0 leaderboard (GPT-5.5) · Tencent
GPT-5.6 Sol15.2LHTB 1.0 leaderboard (GPT-5.6-sol) · Tencent
Kimi K313.0LHTB 1.0 leaderboard (Kimi K3) · Tencent
Sonnet 4.68.7LHTB 1.0 leaderboard (Claude Sonnet 4.6) · Tencent
DeepSeek-V4-Pro6.5LHTB 1.0 leaderboard (DeepSeek V4 Pro) · Tencent
Kimi K2.7 Code6.5LHTB 1.0 leaderboard (Kimi K2.7 Code) · Tencent
MiniMax M36.5LHTB 1.0 leaderboard (MiniMax M3) · Tencent
Doubao Seed 2.1 Pro4.3LHTB 1.0 leaderboard (Doubao Seed 2.1 Pro) · Tencent
Gemini 3.1 Pro Preview4.3LHTB 1.0 leaderboard (Gemini 3.1 Pro) · Tencent
GLM-5.14.3LHTB 1.0 leaderboard (GLM 5.1) · Tencent
GPT-5.3-Codex4.3LHTB 1.0 leaderboard (GPT-5.3 Codex) · Tencent
Qwen3.7-Max (source unspecified)4.3LHTB 1.0 leaderboard (Qwen3.7 Max) · Tencent
GLM-5.22.2LHTB 1.0 leaderboard (GLM 5.2) · Tencent
GPT-5.42.2LHTB 1.0 leaderboard (GPT-5.4) · Tencent
Qwen3.6 Plus (2026-04-02)2.2LHTB 1.0 leaderboard (Qwen3.6 Plus) · Tencent
Grok 4.200.0LHTB 1.0 leaderboard (Grok 4.20) · Tencent
Kimi K2.60.0LHTB 1.0 leaderboard (Kimi K2.6) · Tencent
Loading Atlas data…