56 AI Practitioners Vote Terminal-Bench 4 the Most Trusted Frontier LLM Benchmark
AccBalanced · x · 2026-11-16
Chris Barber polled 56 people in AI on which benchmarks they'd trust most to indicate whether an LLM is frontier-level. Results show coding and agentic benchmarks dominate:
- Terminal-Bench 4 (25 votes): coding, the clear leader
- Terminal-Bench Science (20 votes)
- FrontierCode (13) and FrontierSWE (10): both coding
- Humanity's Last Exam (8): expert knowledge
- Vending-Bench (7): business simulation
- 5 votes each: AutomationBench, DeepSWE, PostTrainBench
- Followed by AA-Briefcase, Agents' Last Exam, FrontierCS 2, OSWorld 2 (4 each); ARC-AGI-3, FrontierMath, Harvey LAB, ProgramBench, SWE-sweep (3 each)
Voters include Alex G. Shaw, Andy Konwinski, Steven Dillmann, Dan Hendrycks and others. Hands-on coding/agent benchmarks clearly outrank traditional knowledge Q&A benchmarks in trusted status.
More from Models
- Claude writing only with Wikipedia sentences passes as fully human-written — andrew_n_carr · 2026-11-16
- Gemini 4 Argon caught gaming the grader: disclaimer first, real views after — sebkrier · 2026-11-16
- Liquid AI d1-2B decision model runs on-device: ~68ms gesture reading, 14 decisions/s, zero tokens — AccBalanced · 2026-11-16
- Anthropic user blasts its 'queuing steering' system as awful — willdepue · 2026-11-16
- Rumor: Google Prepping Strongest Gemini Model Yet, Codenamed Krypton — aziz4ai · 2026-11-16
- Blender watch test: DeepSeek does it for $0.80 vs Claude Fable's $119 — HankYeomans · 2026-11-16