Cracking OpenAI's Code: The Real Story Behind GPT-OSS-20b

OpenAI's GPT-OSS-20b model has been a black box for too long. Despite its impressive scores, no one outside the company has managed to reproduce them independently. Why? OpenAI didn't share the tools or agent harness they used, leaving the AI community in the dark.

Breaking Down the Barriers

Well, that was until some enterprising researchers decided to crack the code themselves. They reverse-engineered the model's in-distribution tools. It's a bold move, but instead of hallucinations, they found strong, confident tool calls. A hallmark of a well-trained model, not blind guesses.

But here's the kicker: they built a native harmony agent harness. In plain English, they found a way to communicate with the model in its own language. This bypassed the often lossy Chat Completions conversion, which can muddle the output.

Numbers Don't Lie

The results? The first independent reproduction of OpenAI's published scores. Achieving 60.4% on SWE Verified HIGH (compared to OpenAI's 60.7%), 53.3% on MEDIUM (vs. 53.2%), and a whopping 91.7% on AIME25 with tools (over OpenAI's 90.4%).

Show me the product, OpenAI! Because these numbers are as close to the original as it gets. If independent researchers can do it, why can't OpenAI be more transparent?

Why It Matters

This isn't just about numbers. It's a wake-up call on AI reproducibility and accountability. If we can't reproduce AI models' results, how can we trust them? And without trust, AI risks becoming vaporware, more hype than substance.

So what's next for OpenAI? They need to open the kimono and let others see under the hood. The community deserves it. The market demands it. And frankly, the integrity of AI progress depends on it.

I'll believe it when I see retention numbers. Until then, let's keep pushing for transparency. Because in AI, seeing isn't just believing, it's everything.