Terminal-Bench 4 and Terminal-Bench Science led Chris Barber’s survey of 56 people in AI about which benchmarks they’d trust to indicate whether a large language model (LLM) is “frontier,” meaning among the most capable models. In results published November 15, 2026, Barber reported 25 net votes for Terminal-Bench 4 and 20 for Terminal-Bench Science. Source
Barber wrote: “The most common comment was that people don't really trust benchmarks at all.” The leaderboard comes with its own raised eyebrow. Source
The leading benchmarks
The totals are net votes, not model scores: Barber says some benchmarks received negative votes. Frontier List supplies the subject categories. Source Source
AutomationBench, DeepSWE and PostTrainBench followed with five net votes each. Source
Read the rules before the victory lap
Barber says participants couldn’t pick a benchmark their company worked on, an index or a human-voted board. Source
He also disclosed that he is a small investor in Abundant, Cognition, Emulated, Fleet, Hillclimb, Latch and Snorkel, and said he did not vote. Source
Barber’s post and Frontier List page don’t specify how participants were recruited, how many votes each could cast or the detailed negative-vote procedure. Neither gives comment counts or underlying responses establishing how many of the 56 respondents expressed distrust. Source Source
The tally doesn’t establish an industry-wide consensus or show which benchmark best identifies frontier models. The poll doesn’t put the benchmarks themselves through an exam.
Comments