UCSB AI Benchmarks
A suite of benchmark leaderboards that measures how the AI models available through UCSB's gateways and inference providers perform in practice — so faculty, researchers, staff, and students can choose the most suitable model and provider for each use case.
Why this benchmark exists
Beyond commercial frontier models, the UCSB community needs AI tools that operate within a safe and transparent environment with strong data-privacy guarantees. Models served through UCSB gateways and inference providers come with assurances that your data is never used for training — and in some cases, is never processed outside university premises.
These leaderboards help you select the right model for specific tasks such as data processing, agentic coding, prototyping, or literature review; understand each model's possibilities and limitations; and track how every model and provider improves over time.
Data privacy
Gateway models guarantee that your prompts and outputs are never used for training.
Transparency
Independent, reproducible evaluations of every model and provider combination.
Over time
Results are refreshed as new evaluation runs complete, so you can track progress.
Current leaderboards
Operational QoS
Operational quality-of-service telemetry that characterizes the end-user experience of each provider’s infrastructure — time to first token, inter-token latency, and generation throughput.
View leaderboard →Scientific codingSciCode
A benchmark that examines how well models generate code to solve real scientific research problems, based on the scientist-curated SciCode benchmark of laboratory problems across 16 disciplines.
View leaderboard →HumanitiesRISE Humanities Data Benchmark
A benchmark that assesses the performance of LLMs on humanities-related tasks.
View leaderboard →