Hi Product Hunt, meet nerfWatch(). When an AI model feels worse, one disappointing answer doesn't tell you much. nerfWatch() gives you a place to compare repeated tests with what people are experiencing. Here's how it works: - Each tracked model takes the same 100 questions daily, across reasoning, logic, reading code, following instructions, and long documents. - Answers are graded by code. Each model is compared with its own first 7 days of tracking, with model settings held fixed. - Community votes sit beside the benchmark. You can vote once per model per day; votes never change the measured score or verdict. - Email alerts report when a model enters the worse-than-baseline state or recovers. - The testing engine is open source, so you can inspect it or run your own checks. Tracking has just started: the initial baselines are still building, so there are no meaningful before-and-after verdicts yet. The dashboard shows "Too soon" while it collects that first week. A few limits matter. These are API tests, not tests of ChatGPT, Claude.ai, or other chat apps. They cover a specific set of tasks, and a score change doesn't reveal why it happened. We also can't measure changes from before tracking began. Take a look at https://nerfwatch.lol and share what you think. Which model should be tracked next, and what kind of test would help you trust the result?

nerfWatch()

Daily AI benchmarks and community votes on model quality

Daily AI benchmarks and community votes on model quality