How we test AI tools

Checked facts

Prices, free plans, Indonesian language support and features are copied from the vendor's official pages. Each fact stores its source URL and the date we checked it, and the page shows both. Facts older than 90 days are marked "needs re-check". A weekly checker also visits every tool: it records the HTTP status, notices redirects to another domain and parked domains, and compares the pricing page text with the last visit. A tool that fails three checks in a row is marked inactive.

Practical tests

For the "AI for X" pages we give several tools the same tasks, with checklists written before the first call. Every call is recorded with the prompt, the model name returned by the API, the date and the cost, including failed calls. In the first run on 25 October 2026 we made 71 image calls that returned an image across image generation and photo editing, graded 130 answers from coding and chat assistants, and reviewed every video clip on four sampled frames. Outputs are shown unedited and labeled as our test.

The NS-ID benchmark

NS-ID measures how well AI models handle Bahasa Indonesia: 60 tasks in six categories (comprehension, formal writing, register, translation, Indonesia-context knowledge, instruction following). Version nsid-v1 tested 20 models from 11 vendors. Exact answers are scored automatically; writing is scored 1 to 5 with a public rubric by two AI judges from vendors other than the tested model's vendor, with a third judge when they disagree. A rerun on two models stayed within plus or minus 3 points. The full rules are in the benchmark methodology.

What we never do

• Accept payment for a test, a score or a ranking position.
• Publish a score without a dated, versioned run behind it, or change one without a public reason in the score change log.
• Show sample or invented results on live pages.