UK AISI and EvalEval Publish Reproducible AI Benchmark Results via Evaluation Cards

The UK AI Security Institute (AISI) has partnered with the EvalEval coalition to release a set of benchmark results through the Evaluation Cards platform.

Why reproducibility matters

As AI models grow larger, the cost and complexity of re‑running evaluations increase, making it hard for researchers to verify published numbers. EvalEval’s Every Eval Ever schema and Evaluation Cards aim to capture not just the final score but also the full configuration, prompts, and compute budget needed to repeat a test.

AISI is sharing transcript‑level details for five major benchmarks—HealthBench, FrontierMath, Humanity's Last Exam, SWE‑Bench Pro, and Terminal‑Bench 2.0—across six frontier models ranging from Claude Opus 4 to GPT‑5.4. The release also includes two cyber‑focused evaluations that use a partially overlapping model set.

Implications for GPU and AI infrastructure

When evaluation protocols and inference‑time compute are disclosed alongside results, engineers can better size GPU clusters for specific workloads and predict performance scaling. Transparent reporting also helps buyers compare hardware options on a common, verifiable basis.

By making the data openly available, AISI and EvalEval hope to close gaps in evaluation science and foster a more reliable ecosystem for model development and procurement.