How to test AI models yourself: our method, and the blind test

How GlanceHall tests coding models without scores: identical prompts, one attempt, fixed recording scripts and a blind test you can run in a minute. Plus what public benchmarks can and cannot show.

By GlanceHall · Updated

The fastest way to test an AI model for your own work is not to read its benchmark scores. It is to give it a task shaped like your work, give a second model the identical task, and look at the two results without knowing which is which. GlanceHall is built around that idea, and this shelf explains how we do it and how you can repeat it.

The method, in five rules:

  1. Same prompt, word for word. Every model receives the identical text; nothing is rephrased per vendor.
  2. One attempt. No follow-up messages, no reruns for a better result. Only an infrastructure failure, such as a rate limit, earns a retry, and it is logged.
  3. Fixed content. Product names, copy, prices and data are set in the brief, so differences come from execution.
  4. Identical recording. The same script scrolls, clicks and plays every build at the same speed; games run on a virtual clock so a slow machine cannot make one look choppy.
  5. No scores. Facts are published, judgement is left to you.

Try rule five on yourself: four landing pages, names hidden, pick before you reveal.

Take the blind test

Public benchmarks still have a place, and the pages below explain what they measure and what they cannot. The full procedure, including the exact commands and the 150-minute limit, is on the methodology page.

3 pages in Testing & method

Testing & methodGPT-6 benchmarks vs what it built: the scores, and the builds you can checkGPT-6 Astra's benchmark scores from OpenAI and from public leaderboards, set against what GPT-6 Astra and GPT-6.1 Sol actually built on identical briefs. Where the numbers hold, and what they miss.Testing & methodAI comparison, done visually: four models, eight briefs, side by sideCompare AI models by watching what they build. Pick a brief, play all four builds at once, pit any two against each other, or hide the names and take the blind test. No scores, just evidence.Testing & methodAI coding benchmarks explained: SWE-bench, Terminal-Bench, LiveCodeBench, arenas, and what they can't showWhat the major AI coding benchmarks actually measure, how to read their scores, where vendors' numbers and public leaderboards disagree, and the things no coding benchmark can tell you.