AI benchmarks
These AI benchmarks are tests we run ourselves with API keys we pay for. We publish how each test works, the cost of every run, and every change to a score. Payment and sponsorship never change who is tested or what a model scores.
Short answer
One benchmark suite is live: NS-ID v1, a Bahasa Indonesia test of 20 models run on October 25, 2026. Speed tests run weekly on direct vendor APIs. An English suite is planned and has no scores yet.
60 tasks in six categories, 20% published as a sample. Writing tasks are scored by two AI judges from other vendors.
NS-ID scores next to official prices, real cost per 1,000 tasks and measured speed.
Time to first token and tokens per second on each vendor's own API, median of repeated streamed calls.
How models are picked, how tasks are scored, the monthly budget, independence rules and the schedule.
Every published score and every change to one, with the reason.
Planned. We will not show English scores until we have run our own English tests.