16.6pp above Opus 4.6 (65.4%). Terminal-based coding tasks.

Tied for #1 on Terminal-Bench 2.0 leaderboard. Agent framework matters: same model scores differently with different agents.

Tied for #1 on Terminal-Bench 2.0 leaderboard with GPT-5.4. ForgeCode agent framework achieves best results for both models.

#3 on Terminal-Bench 2.0 leaderboard. TongAgents framework outperforms ForgeCode (78.4%) for this model.

Terminal-specific model variant. Leads Terminal-Bench over GPT-5.4 (75.1%).

#1 open-source on Terminal-Bench. Behind GPT-5.4 (75.1) and Mythos (82.0) overall.

Beats Claude Opus 4.6 (59.3%) on terminal operations

Terminal/CLI tasks. Solid for a smaller, faster model.

Default thinking configuration. Measures terminal/CLI task completion.

Official blog. Matches GPT-5.3-Codex on SWE-Pro (56.22%). Self-evolution architecture.

Terminus-2 framework. With verified subset (fixed ambiguous instructions): 60.7%. Strong for an open model.

Open-weight MoE with 11B active params on terminal coding tasks.

Terminal/coding benchmark. Behind Qwen 3.6 Plus (61.6) and Claude Opus 4.6.

Terminal tasks. Budget-tier model optimized for speed over depth.