Here's something you might have noticed if you've used OpenAI's o1 or o3 models for anything involving contracts, financial analysis, or regulatory compliance: the improvement over GPT-4o isn't that it knows more stuff. It's that it thinks differently.
Ask GPT-4o to track fifteen conditions across a licensing agreement and you'll watch it lose the thread somewhere around condition nine. Ask o1 the same question and it will methodically work through each constraint, correctly flagging where clause 12 creates an exception to clause 4, which itself only applies when the conditions in clause 7 are met.
That kind of structured reasoning doesn't emerge from throwing more data at a model. It requires teaching the system how experts actually think through layered problems. And that's where Project Strawberry came in.
From Mercor to OpenAI
David Nagtzaam's involvement with OpenAI's reasoning initiative grew directly from his work on Mercor's Project Argentum and APEX frameworks. Those projects focused on encoding domain expertise into training rubrics for frontier models. The OpenAI collaboration extended that work into a more targeted application: improving how models handle the kind of conditional, multi-step logic that defines professional knowledge work.
The project, internally codenamed Strawberry before the o1 release (May 2024), brought together domain experts who could demonstrate reasoning patterns that earlier models consistently failed to replicate. Not answering questions correctly in isolation, but maintaining coherent logic across extended chains of dependencies.
What Actually Changed
The difference shows up in practical applications. Previous models would handle individual steps competently but struggle to thread constraints across an entire problem space. A marketing funnel with budget caps, audience segments, channel restrictions, and sequential dependencies would produce outputs that worked locally but contradicted themselves globally.
The o1 architecture handles these scenarios differently. It maintains state across reasoning steps, tracks exceptions to rules, and catches when an earlier assumption invalidates a later conclusion. For professionals working in regulated industries or complex operational environments, this represents a genuine capability shift rather than incremental improvement.
Legal contracts. Compliance frameworks. Financial models with nested assumptions. Strategic plans with interdependent initiatives. These are domains where "if X, then Y, except when Z, unless condition W applies" describes everyday work. Training models to navigate this logic required experts who live in these environments and could articulate the reasoning patterns that become invisible through practice.
The Expert Layer in AI Development
OpenAI's approach to improving reasoning capabilities reflects a broader recognition in frontier AI development: synthetic data and scale hit limits. Pushing models beyond those limits requires human expertise that can demonstrate what better reasoning actually looks like.
For David, the OpenAI work reinforced insights from the Mercor collaboration. "The challenge isn't teaching models what to conclude," he observes. "It's teaching them how to hold complexity without dropping pieces. That's a skill that comes from years of professional practice, and it's surprisingly hard to transfer."
The collaboration positions Decode at the intersection of AI capability development and practical business application. As organizations adopt reasoning-focused models for high-stakes work, understanding both what these systems can do and how they learned to do it becomes a meaningful differentiator.
Ready to put advanced AI reasoning to work in your organization? Decode helps companies implement AI solutions that match the complexity of real business problems. Get in touch to explore what's possible.