Spring Prompt: Independent AI Model Benchmarks for Real Business Work
How we built a research site and testing tool that show which AI model suits a job, using evidence rather than launch claims
Tags: AI evaluation, LLM benchmarks, model selection, prompt testing, AI product
New AI models arrive every week, and each one comes with charts saying it's the best yet. For a business choosing a model, those charts rarely answer the real question: will it do our job well, reliably and at a sensible cost
We kept running into that question with our own clients, so we built Spring Prompt.
What we built
Spring Prompt is an independent research site and testing tool for AI models. It has three main parts.
Benchmarks based on real work. Most public tests measure general intelligence. We wanted to know how models handle the jobs businesses give them, so we built our own LLM benchmarks for tasks like writing product content and making slide decks, and we publish real model outputs alongside the scores.
One view of the market. The AI model leaderboard combines our five benchmarks with eight licensed sources into one score per model, and every part of that score is shown. We also track API prices and new model releases across 21 providers.
Testing your own prompts. A leaderboard can only tell you so much, because your prompts aren't ours. The Prompt test runs a prompt on realistic inputs across several models, checks every answer, then rewrites the prompt from its failures and tests it again. It also works directly from coding agents, so a developer can test the prompts in their own codebase.
Why it matters for our clients
Spring Prompt is how we stay honest about AI. When we recommend a model to a client, the recommendation comes from tests we can show, and when a cheaper or faster model does the job just as well, we can prove that too.
It also keeps us close to the technology. Testing every major launch means we see where models improve and where they still fall short, and that feeds directly into the systems we build.
