LiveBench

LiveBench was built to answer a specific worry about every other benchmark on this list: that a widely used test eventually leaks into a model’s training data, so a high score reflects memorisation rather than ability. A team spanning Abacus.AI, NYU and the University of Maryland — including Tom Goldstein and Yann LeCun among its authors — published it in June 2024 with a fix: draw questions from sources too recent to have been trained on, such as that month’s arXiv papers, news articles and maths competitions, refresh the set on a rolling schedule, and grade every answer against an objective ground truth rather than a human or LLM judge’s opinion.

That design also sidesteps a second problem, the one that dogs arenas and LLM-judged benchmarks alike: a judge model can be gamed by style rather than substance. LiveBench’s 23 tasks, spread across reasoning, coding, agentic coding, mathematics, data analysis, language and instruction following, are checked against fixed answers instead. The tradeoff is that the whole question set is periodically replaced — the site had moved from its original 2024-06-24 release to a 2026-06-25 edition by the time of writing — which keeps scores comparison-resistant across versions in the way a static benchmark’s are not.

On the current release, Anthropic’s Claude Fable 5 led with an overall score of 83.0, narrowly ahead of OpenAI’s GPT-5.6 Sol at 81.0, with cost-per-successful-task published alongside every score so a reader can weigh capability against price directly. Because each refresh changes the underlying questions, LiveBench functions less as a fixed yardstick than as a standing methodology for building one — a response to benchmark contamination rather than a single benchmark in the usual sense.