Terminal-Bench 2.1

Name: Terminal-Bench 2.1
Creator: Vals AI
Keywords: Coding

AI benchmark · Coding

Latest from Vals AI on X

Do agent teams pay off? We ran GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench, alone and as teams, at medium and max reasoning effort. Teams cost 1.8× to 5.1× more, but only one of the four team-vs-single comparisons produced a statistically significant improvement: Sol at medium effort, which gained 7.3 points.
— ValsAI (@ValsAI) Oct 9, 2026
What happens when a teenager pushes back on an AI’s safety guardrails? 1 in 5 teens now talk to AI chatbots about mental health problems, but these chatbots can make mistakes. Some of the most dangerous mistakes happen later in conversations, after the model initially responds appropriately. We built the SAFE-Teen benchmark to test how models handle serious conversations with adolescents.
— ValsAI (@ValsAI) Oct 8, 2026
We tested Mercury Decide, a new model from @_inception_ai, on the same benchmark we used for TypeSafe's Jev. It matched frontier accuracy on claim verification at the lowest cost we have measured so far.
— ValsAI (@ValsAI) Oct 8, 2026

State-of-the-art set of difficult terminal-based tasks

Its domains, and the nearest entries sharing them. Click any node to open its page.

Next: browse all benchmarks. Entry from the RL Research daily scrape of public sources, 2026-10-10 snapshot.